Repository navigation
minibrain 0.1.1
Install
- Installer:
curl -LsSf https://github.com/getminibrain/minibrain/releases/download/v0.1.1/install.sh | MINIBRAIN_VERSION=0.1.1 sh - Wheel:
uv tool install --python 3.12 https://github.com/getminibrain/minibrain/releases/download/v0.1.1/minibrain-0.1.1-py3-none-any.whl - Docker, CPU:
docker run -p 127.0.0.1:8000:8000 -v minibrain-data:/data ghcr.io/getminibrain/minibrain:0.1.1 - Docker, NVIDIA CUDA 13.2:
ghcr.io/getminibrain/minibrain:0.1.1-cuda
Verify
Every file has a Sigstore bundle (<file>.sigstore.json) and every image a signature, signed keyless by this
release's workflow run. With cosign:
- Files:
cosign verify-blob --bundle SHA256SUMS.sigstore.json --certificate-identity https://github.com/getminibrain/minibrain/.github/workflows/release.yml@refs/tags/v0.1.1 --certificate-oidc-issuer https://token.actions.githubusercontent.com SHA256SUMS, then
sha256sum --check --ignore-missing SHA256SUMS - Images:
cosign verify --certificate-identity https://github.com/getminibrain/minibrain/.github/workflows/release.yml@refs/tags/v0.1.1 --certificate-oidc-issuer https://token.actions.githubusercontent.com ghcr.io/getminibrain/minibrain:0.1.1 - Provenance:
gh attestation verify <file or oci://ghcr.io/getminibrain/minibrain:0.1.1> --repo getminibrain/minibrain - Offline and air-gapped: docs/install.md
Guide: docs/install.md
Changes since v0.1.0
Features
- api: creating a dataset takes its description and tags, so a bad tag refuses the create instead of leaving a half-made dataset (9f4d686)
- data: focused dataset editor page: a compact title bar over the query and the preview side by side at full height with a resizable divider; the SQL editor fills its pane (6ee8ab5)
- data: DuckDB function and keyword completion with signatures, an Examples menu of query recipes, and clickable columns in the tables box of the SQL editor (4c3f4c3)
- data: Quality tab in the dataset editor with a verdict line, each finding's rows and examples, Show rows and one-click fixes; Clean up options saved with picked columns; personal data in the quick checks and a random sample of large results (a220775)
- infrastructure: clearer machines table: one Running column for models and jobs, absolute and per-GPU memory, quiet pool rows, a totals row, no machine expansion (7a30a6f)
- data: Run button (⌘↵) in the dataset editor; the preview runs as you type only on sources under a million rows, previews take at most a quarter of memory and half the cores, one preview per person (265f9e7)
- audit: keep every hash-chain break until an admin acknowledges it with a note (35b4344)
- infrastructure: pool row menu opens Pool settings, pool rows show the description; drop the unused PoolMenu (1ed493a)
- infrastructure: optional pool description, set with what the pool is used for in Pool settings and shown in pool cards and pickers (8224b41)
- infrastructure: one machines table with a Status column, model chips by name that fit the column, plain job progress and quiet pool rows; drop the temporary layout switcher and layouts B, C and D (2cc49d2)
- audit: verify the hash chain hourly and nightly, alert on breaks and SIEM lag, record audit views and exports (d9566c8)
- infrastructure: machines table columns resize, fit their content and persist; four layouts to compare behind a temporary switcher (refined, flat with filters, memory map, list and detail) (8bcbb7e)
- data: Pick columns lists each column with type, sample values and distinct count, suggests the text and answer, combines sources by name and skips rows with no answer; the SQL editor formats, folds, runs a selection and shows column values (bf88588)
- data: one Edit dataset page for name, description, tags and query, shared with New dataset (01ede9c)
- data: build a dataset's query on its own page with a live preview (488dd7e)
- models: Model, Quality and Requests headers sort the list with the shared obp-ui SortButton; drop the unused local SortButton (003b5ed)
- infrastructure: machines as one refined table with comfortable rows (5b70a84)
- compute: each pool is used for any work, serving or training (1367ace)
- web: one error layout everywhere: page failures fill the content area, section failures stay in their section, inline errors keep a Details disclosure with status, path and request id; retry refetches in place (14f0e48)
- data: open saved dataset versions from a header picker (8854df9)
- data: Sources has Add source and each source group its own primary action (Upload file, Add folder, Add S3 path, Add database table, Add lakehouse table), blocked with the reason when not set up (ac9e381)
- data: the Datasets page has New dataset as its primary action; passed checks open in place as a ticked list with each detail (6d33974)
- data: the catalogue tree remembers which branches you collapsed (be415ab)
- feedback: review queue with status tabs, word diffs, one-at-a-time review and a train nudge per model (1f835f7)
- playground: Recent groups chats by day, one line each with the model's avatar, title and age; Rename and Delete sit in a menu shown on hover (abd1f2d)
- data: Add moves to a + button beside the catalogue search, leaving one primary per screen (0d78f38)
- data: catalogue rows are real links, so middle-click and the browser menu open them in a new tab (87ce38d)
- models: Add model greys out datasets whose latest version has too few rows, with the reason (f2f438c)
- playground: the message box takes focus when the Playground opens; preset buttons carry icons (86008f5)
- data: previews read First 20 of 750 rows with Show more, up to 500 rows (2929643)
- data: train on as few as 10 rows; under 100 the check warns that scores will be rough instead of blocking (81b5bc1)
- data: path URLs for the Data page, saved-version stats, too-few-rows block and Save full version (0bb0590)
- data: the About column resizes like the catalogue tree, by drag or arrow keys, and remembers its width (4d7d66f)
- training: Version is its own column again, sortable on the server (50114e7)
- data: query results carry DuckDB column types; preview and SQL result tables sort numbers, dates and yes/no by type with empty cells last (16bcb29)
- data: folder page keeps New dataset as its one primary action; Rename, Move to… and Delete move into a … menu (57541d2)
- machines: group machines by pool or by model; machine rows expand to the models and jobs they run, and the model view lists each deployed model's replicas, including models with none running (d1e6d05)
- tables: every table sorts by its columns; Jobs, Training, Requests and Audit log sort on the server with keyset paging and keep the order in the URL (ce6b6df)
- data: item header keeps one primary action and a … menu; About narrows to 320px under 1440; list grids fill the page height (378a8dd)
- infrastructure: Storage in two columns: disks on main beside other machines, usage groups that open in place beside a compact Free up space list; Stores moves to Setup (cd61ae8)
- model-overview: a one-line next step until the model has first been live (0478946)
- model-page: problems and versions side by side, one-line versions, visible runs (d955896)
- model-card: the card reads as a document with a contents list, folds long lists, has one Download menu, says Live/Ready to deploy/Not deployed, and renders before data (fd028ab)
- exports: export a version as a merged full model in safetensors (a0b8af6)
- data: item detail in three columns, About beside a Records, Table and Example preview; Schema with plain type words; compact checks, label balance and versions; query edited in a dialog; nested source values come back as JSON (40d2ff7)
- extensions: minibrain extensions install fetches the default set from DuckDB's repository when no folder or mirror is given; dev setup installs them (503960e)
- storage: What's using it by base, adapter, dataset, upload, job and database; Free up space as suggestions; job log retention; machine cache size (e3f482f)
- attention: Needs attention on the model Overview, scoped to the model, and on Home across everything (6bd2109)
- model-page: Versions panel and usage strip on the Overview, Activity tab, four-chart Monitoring with recent requests, Versions grid fits 1280px (590ce6e)
- evaluations: one row per model with live and latest score, ready to deploy first then regressions; Where it fails gets By label and an honest empty state; hollow dots for versions that miss the deploy rule (618d0b1)
- infrastructure: problems endpoint and Needs attention; tabs become Machines, Storage and Setup (98eac96)
- job-page: Progress, Inputs and Result sections with the version page's verdict line; the Jobs list shows Pool and Duration (28e34b4)
- version-page: fixed layout of summary, training and evaluation, a version picker, a version note and "deploy rule" wording (dd04c09)
- cost: price scores and traffic per 1M tokens, the unit APIs price in; Cost is off by default in the Models list (7b3be4e)
- training: the Training list shows score change, base and data, GPU-hours and ETA; list times share one When (db22e59)
- machines: one join token per pool, rotate to cut off the old one, and approve machines that join without a token (e24908c)
- models: the models table loads with skeleton rows shaped like its columns, fading down the page (fe8f163)
- ui: every status dot is obp-ui StatusDot, with a progress arc, ping while active and a dashed ring while queued (4a93ff5)
- infrastructure: machines report live GPU memory, load and disk with each heartbeat; admins drain, undrain and remove them; pools show the room left (8e1b4ac)
- infrastructure: Add machine creates a pool's join token, shows the join commands and waits for the machine; Machines lists join tokens with revoke (21944c5)
- security: per-pool join tokens replace the shared worker token for machines outside the local pool (9664dd0)
- training: the worker records GPU-hours (CPU-hours on CPU) per training run; the version page shows them (163e477)
- versions: a run records the version its settings started from; lineage, training summary, model card and AI-BOM show it (aa50fc5)
- versions: version page crumbs end at Versions, step arrows to the previous and next version, queued runs read Queued (dde4f18)
- infrastructure: Machines is one filterable table grouped by pool with a detail sheet per machine (f5b840d)
- compute: machines report their recent jobs, CUDA and NVIDIA driver versions (6c3cbfc)
- account: Your account page to rename yourself, change your password and sign out other sessions (6a99840)
- infrastructure: Storage shows minibrain's share, mount points, owners of large items and what could be freed (e358147)
- models: the version page is one fixed skeleton in every state (a90a2f0)
- data: take a seeded random sample when a query returns more rows than the limit (5c3920e)
- training: loss and learning rate share one figure with phases, markers, smoothing and a verdict (0176cbe)
- settings: every settings screen uses one two-column SettingsSection layout (9bc9810)
- infrastructure: fleet board, setup diagram, fixes for each issue and per-disk storage (5386daf)
- training: learning-rate schedule, LoRA dropout and weight decay are training parameters (dcdcbde)
- web: links on create pages open in a new tab, marked with an icon, so the form keeps its state (04b389f)
- training: learning-rate warmup is a training parameter, 5% of the steps by default (3047df3)
- training: compare a run's loss with any trained version, the live one by default (11473b9)
- models: a version's logs live only on its job page (926c7f3)
- models: starting a training run opens its version page, which follows the run live (e64bdaa)
- web: crumbs show only parent pages, so top-level pages have none and the sidebar group label is gone (8491938)
- playground: model picker shows custom-model avatars and its hint; typing after opening goes to search (8de9952)
- models: Connect dialog is three steps on a rail (key, code, check), endpoint values on one strip, real logos for every harness and code on a dark surface (96a8cff)
- models: Add model's Dataset menu uploads a new dataset in a dialog and adds demo data in place, picking the result (2eccc44)
- models: Add model has an Output row (Labels, JSON or Free text) suggested from the dataset's answers, sent as the contract so v0 is scored the right way (7976e63)
- models: Deploy dialogs and Settings › Deployment set replicas within the pool's room; list and bulk Deploy pick a pool (6252755)
- serving: moving a live deployment to another pool hands over once the new pool is ready; replicas are bounded by the pool (f1f0151)
- models: a model's dataset is locked while a training or scoring job of it is queued or running (01d191e)
- worker: a worker whose name resolves to loopback advertises the address that reaches main, or refuses to start with the --advertise fix (45886ac)
- security: workers hand their engines a per-pool engine token instead of the gateway token (422e643)
- data: large demo data: demo_data/build.py --large builds DBpedia-14 (630,000 rows) and Amazon reviews (4,000,000 rows), listed under Demo data once built (6950964)
- settings: Audit integrity and Audit retention move to Settings › Security; the Audit log page only browses, with one status line linking to them (cc53053)
- home: getting-started checklist follows the core journey from real state, hides when done or dismissed (49fc372)
- training: Train page validates on click, marks and focuses the first bad field; Instructions grow to ten lines, capped at 8,000 characters with a count (cd5c19e)
- evaluation: What it's tested on and Test examples are one section; Remove shows on a row's hover or focus (18c1adc)
- models: list rows drop their buttons for the … menu (Try it, Connect, Deploy, Undeploy); Quality is one line: score and gain over prompt-only (e22cc48)
- playground: model picker moves into the chat composer; header keeps New chat and folds the rest under a menu (6075b1c)
- models: Quality cell shows the score on a 0-100 bar with the prompt-only mark, the gain over it and a better candidate; hover lists live, best and prompt-only (d3c9782)
- models: Deploy dialogs get a Pool row with each pool's fit; header, Overview and Models list name the pool (8cf701b)
- serving: deployments run on a chosen compute pool; workers there start and stop their engines (204b9a0)
- web: one page header with a crumb row rooted at the sidebar group on every page (b44e487)
- settings: every settings page gets a one-line caption; Connections drops its fake -mt-2 caption (ad0123c)
- models: Usage shows requests, p95, errors and cost side by side over 24h by default, with a zero line when quiet; metric checkboxes; Activity gets a caption (2291154)
- models: Request content moves to model Settings; Monitoring's request log links to it; Safety model hidden from the UI for v1 (03f70be)
- models: API Connect picks a language from a dropdown with logos; Deploy shows the model's avatar and API name (3d1cced)
- models: API Connect creates a key in place, fills it into the code, sends a test request and shows when the first request arrives (afb5ec4)
- models: stepped API Connect dialog with endpoint, key, call and feedback; API keys tab drops Endpoint and Feedback (b4888d8)
- models: Add model lists every pool with its memory bar and whether the model runs and trains there; disk gets its own section (6b0dc61)
- samples: demo datasets are named Demo travel intents, Demo Banking77 and Demo CLINC150 (c7a1bc6)
- models: Add model summary shows one Pool with memory and disk bars measured on the pool it runs on; notices move to the bottom of the form; create panels share one height (e9d619e)
- models: Evaluation tab with scores by version, where it fails, what it's tested on, test examples and how it's scored, moved out of Settings (02253f0)
- web: one ErrorState for not-found, forbidden, failed-load, unknown-route and crash pages; inline errors get an icon and Retry (2ebc485)
- ui: name every place by pool, with a shared PoolLink hover card (0294cef)
- versions: v0 cannot be deleted; it is the baseline every version is scored against (fd1cea1)
- models: Add model checks on Create: errors appear next to their fields after the first click, and the button stays enabled until then (fd47128)
- training: pool name links to Infrastructure with a hover card, and the preset dropdown shows icons and descriptions (380be32)
- models: Train page with pool compute, a collapsible Training card and inline instructions (a4a192c)
- api: a training run can target a pool and only that pool's workers claim it (ae2fa86)
- models: Train page on the shared create-page layout (e0431d2)
- models: create-page actions under a divider with Cancel then the primary at the right (08d8612)
- models: Add model on a shared create-page layout with a required dataset (0dcd8b5)
- api: a custom model needs a dataset when it is created (7b240e6)
- models: merge the Training tab into Versions, run details on the version page, Instructions only on the Train page (4fe30f0)
- versions: save new instructions with the version that uses them (6f1321e)
- models: Settings tab with re-scoring, Train · Deploy · Use header, deployment status by the name (83a6046)
- training: Train page replaces the Train dialog, with a sticky run summary, two-column parameters and Train again on versions (f0e78b6)
- proto: model Settings tab replaces the header menu, re-score on grader or judge change, Scored-by line, dataset required on Add model, Train dialog starts from the last run (15d3428)
- training: train dialog in two titled panes, General and Parameters, with presets above always-visible parameters and the memory gauge under Compute (e7c39a4)
- training: train dialog with a dropdown per row, dataset version first, and the memory gauge and estimate in the footer (9b6bcd4)
- training: train dialog with data first, Fast, Balanced, Best and Custom, and a Parameters column that unlocks for Custom, with hints in tooltips (13c98f4)
- models: a Data column with each model's dataset, the version it trained on and a newer one when available (ebd2561)
- audit: area, target and time-range filters, actor names and kinds, target links, result column, export menu and a side panel for event details (c089433)
- proto: train dialog without tabs, dropdowns for dynamic lists, Custom quality in three layouts; base and dataset dropdowns on Add model (df48aeb)
- proto: bind a custom model to one dataset at creation or first training, preselect the best machine, keep jobs-only pools out of deploy (4a736d6)
- proto: fixed-height train and deploy dialogs with General and Advanced tabs and an always-visible estimate (888d196)
- proto: static Deploy button, Use menu with Playground and API Connect, Edit and Undeploy on the deployment (6eb34a9)
- proto: deploy and undeploy as jobs with rollout steps and logs, and replica logs on Monitoring (d58e3b8)
- proto: three train dialog designs to compare (590a834)
- api: label the two API references, group /api routes into sections, hide /internal (aaf752d)
- keys: all-model API keys and an API keys page under Operate (166e444)
- models: deleted models' API names stay reserved and never reach another model (3476f51)
- gateway: OpenAI-compatible /v1/completions (d533d6b)
- proto: header actions, a deploy dialog and an infrastructure view of what runs where (dde6f84)
- models: API name with a copy button under the model title (6c6b61a)
- models: clearer model activity with names, grouping and day captions (5f49584)
- models: instructions and how it's judged move to the Training tab (d81ad17)
- models: Try it label, usage chart dropdowns, no base in header subtitle, base-size empty and total captions (7205d97)
- models: list Reset to default, plain group headings, full-cell chevron and menu hit areas, no help cursor (9939601)
- models: avatars and generated names for custom models (6205555)
- playground: message input matches the kit's dock composer (87ce2e8)
- jobs: delete jobs singly or in bulk (4ff8841)
- models: base picker with Catalogue and Your imports, and a clearer import form (55d5900)
- proto: avatar picker in three steps: style, shape, colour (05f10aa)
- proto: avatar styles as four cards with fixed sample icons instead of tabs (fd35f6e)
- models: the list's 7-day requests, p95, errors and device time come from the model page's traffic statistics (a385c8d)
- web: models list with groups, view options, versions and variants, selection and a quick look (576e1b2)
- api: model rows carry p95, errors, live replicas and base variant facts in a fixed number of queries (c0fe34e)
- web: model page from the approved prototype (2b30044)
- web: shared line chart on StepChart with validated series, latency and memory chart tokens (4d53c94)
- api: per-model traffic, activity and status line; named keys; real planner times per machine (679e849)
- web: sidebar by role with Monitoring and Jobs, Training runs and Jobs pages, canonical job pages with settings, loss and learning rate (4d0d395)
- api: jobs carry method, adapter, score and base; a job's setup; method filter and kind facets (6d54c79)
- proto: monitoring tab and page share the overview's usage chart, with latency, errors, versions and callers (9867b99)
- proto: a train dialog with quality presets, data and compute pickers and grouped advanced settings (3c61319)
- proto: avatar picker with a large preview, hover and compare, and surprise me (5b70323)
- proto: model details as six key facts with all details behind an expander (9bf9c12)
- proto: a much better avatar picker and name generator (4bf2c8a)
- playground: the conversation uses the obp-ui chat set, with a copy button on each answer and reasoning that closes once the answer starts (f8172f6)
- proto: model overview with details under the chart, and Use it and Activity on the right (e7c5de7)
- proto: one canonical URL per job, training jobs under Training and the rest under Jobs, with a context line (16abbd8)
- proto: compact model header, priority-driven primary, a fact-sheet column, a status column for versions, Make vN live and Export (86f2964)
- proto: models page header with title and subtitle, then search, filters, view options and Add model in one row (5d4c32b)
- proto: three models page headers to compare, and custom and base groups (0ae4734)
- proto: versions as a report card, training on the job, model training as its jobs (7bb27a3)
- proto: one view options panel for the models list; no grouping or delete-models button (800861b)
- proto: versions as commits with a live pointer, home inbox and recipes, command palette, one model page (973669f)
- proto: models list with child-only selection, base variants, replicas, quick look, worth-it columns and autopilot (25a0fd8)
- proto: proposed structure at /proto/flow: models list with nested versions, version page, compare and the full journey (69ca67c)
- models: cancel running jobs from the delete dialogs, then delete once they stop (28c4918)
- proto: framed avatar variants without backgrounds on /proto/model-icons (45ec8ec)
- proto: DiceBear avatar styles and a picker mock on /proto/model-icons (652ae23)
- models: switch a custom model's base without touching live traffic; Train off with the reason on serve-only bases (645acdd)
- proto: twenty custom model icon designs at /proto/model-icons (aa95747)
- models: duplicate a custom model with its settings, data and live version; base picker lists base models only (147bb1d)
- models: one model page at /models/$name for both kinds, with overview, versions timeline, data, API and files (77877cb)
- web: deleted models show as name (deleted) in jobs, requests and overview (bbeacef)
- registry: jobs and request logs outlive a deleted model (6c70277)
- web: a job page at /jobs/$id for every job kind, outcome first, with loss and learning-rate charts and the job log (882811a)
- jobs: structured job events with levels, per-step training metrics with held-out loss, and a job metrics API (892edfb)
- training: tick jobs on Training and Monitor › Jobs to cancel or retry them together (ad215d1)
- playground: Recent chats saved per person and encrypted when request content is kept, and Save as test example into the next version's held-out rows (b25b480)
- web: every code snippet is a highlighted CodeBlock, and API and View code snippets are CodeTabs (ad681db)
- web: model pickers and Model filters show each model's logo beside its name (65ad64c)
- web: a full Playground with settings, presets, streaming, reasoning, tool calls and its code (f486210)
- gateway: a safety model checks requests and answers, per install, model or request, with refusals audited (dcdb5c3)
- gateway: reasoning off or at an effort, returned apart from the answer, and tokens streamed from every engine (6b92651)
- gateway: tool calling through each model family's chat template, with SGLang and vLLM parsers (86951b9)
- gateway: a request's response_format only narrows a model whose contract is free text (1dc8367)
- gateway: sampling parameters on every engine, max_tokens up to the context limit, and a vLLM engine (7b634bf)
- models: Add model asks only for a base and a name; every model starts as a Chat assistant and narrows later (9f88a93)
- web: How it's judged and Instructions on the model page, the answer review, the judge model setting and a conversational Playground (a47b7f6)
- engine: General, shown as Chat assistant: free text over turns, chat JSONL and question and answer tables, a judge model setting with reasons, and the answer review (240a0c6)
- data: one Add menu in the Data header (94c8ef9)
- web: Add model with Classify and Custom, Start from a recipe and free default names; grader scores, the promotion decision and the Classify view; the Task section (7aa8a93)
- engine: every model runs on the task engine: chat-example dataset versions, version 0 at creation, numbered additive contracts, graded scores, judge jobs and the promotion rule (migration 0041) (316f26a)
- engine: contracts, chat examples, graders with a cached judge, the promotion rule, preset files and the recipe format with Ticket routing (1b9266c)
- data: drop a file on Data → Uploads, through the same duplicate and Replace or Keep both flow, with obp-ui's DropZone (628d81b)
- train: plan each run's memory: LoRA or QLoRA, gradient checkpointing, 8-bit optimisers and a fitted sequence length (79336d9)
- api: an air-gapped API reference with a try-it client at /developers, and FastAPI's public docs off (69169ba)
- train: pick CUDA, ROCm, XPU, MPS or CPU, train on the Apple GPU one example per micro-batch, and fall back to the CPU when it fails or stalls (e54150b)
- data: one source per file, Replace or Keep both for a taken name, unique dataset and source names (17839c8)
- models: the disk gauge and the model page name the store; Storage copies bases to the bucket (9e1f859)
- models: base model weights follow Storage, with the bucket as the main copy (470d8d5)
- config: provision the whole install from minibrain.yaml, locked in the UI (03ee7d2)
- settings: search every setting from the settings rail, with links that land on the row (58c93bf)
- models: Explore in Discover link in the base picker header, opening the list in a new tab (e0b4970)
- models: each catalogue row in the base picker opens its Discover page in a new tab (fd2c1e4)
- models: formats say Full precision or Quantized with their scheme, filterable on Discover (7ce63a0)
- proto: ten tab designs for the base model picker at /proto/tabs (c09ee61)
- models: Import tab starts from three source cards and shows what it found (12ca17f)
- models: base picker tabs sit left with one Filters menu on the right (972581c)
- models: the licence checkbox turns red with its message when it is missing (389ec47)
- data: a dataset saves its rows as version 1 and each update is a new version (a1a716c)
- data: select sources and make one dataset from them (ecf21a7)
- data: select several datasets and folders to move or delete them together (520130a)
- models: a base model's API section offers curl, Python, JavaScript and client configs (dceebe0)
- models: delete dialog lists every model and version it removes (3a51caf)
- catalog: context length and release date for every model (851eb24)
- models: My Models groups your models above base models (95b8eb5)
- models: Import tab is one field with clickable examples and a live file checklist (98d5765)
- models: Try with sample data downloads Qwen3 0.6B when it is missing (f0a779e)
- models: show the memory a model needs and filter on Runs here across compute pools (7e36abd)
- models: import inspection returns a checklist of what the folder has and lacks (a0bdd54)
- accounts: Invite person, with an icon, instead of Add person (5e0dbdb)
- models: empty Models page explains the three steps and offers Add model, the catalogue or the sample (1417c17)
- auth: sessions last a working day idle and a week in all, the sign-in page says why you were signed out, and dev has no idle timeout (0cc0e1d)
- web: /proto/type compares the kit's type pairings; obp-ui with Mona Sans for Cobalt (d84f98c)
- models: Add model is always clickable and marks each problem inline; names need a letter or number and conflicts name what you typed (a41995c)
- models: delete, undeploy and base file removal run in the background with Deleting and Stopping states, a cascade dialog and toasts (2ab9c7d)
- models: base model picker shows the memory each model needs and one Runs here checkbox across compute pools; no fit filter or fit badges (9e60639)
- models: disk and memory gauges on Add model show used, this model and free space, with exact sizes (d45a63f)
- models: licence at the end of the Add model form, disk and format notes in the side panel, and a toast when adding fails (acd5c91)
- models: tags help tip laid out with key:value examples in two columns (873c771)
- models: the too-big warning sits in the Add model side panel under Memory (e3cc19e)
- models: too-big models get one warning line with Where to run it and Pick a smaller one; a single format still shows its radio (3d21f0b)
- models: Add model form with taller description, tag help in an info tip, quieter optional labels and 32px between fields (9ff84da)
- models: Add model summary panel with the model, format cards and disk and memory meters (2813852)
- data: delete a dataset that models use; they keep their versions and snapshots and need a new dataset to train again (750aeae)
- web: ten cog brains and ten simpler brains on /proto/logos (0e6856c)
- web: ten simple brain marks on /proto/logos (308c81f)
- web: ten versions each of Branches, Waves and Rainbow on /proto/logos (c57f435)
- web: ten marks on infinity and abundance on /proto/logos, five detailed and five bold (ea6532f)
- web: five new original marks on /proto/logos; the knot is gone, it was Apple Intelligence's mark (6e8e3f7)
Fixes
- release: the PyPI summary and image descriptions say models of any size, not small task-specific models (e5f1b0f)
- image: Python 3.12.16, the latest security release, and uv 0.13.7, which can install it (844e184)
- deps: joserfc 1.7.6 and OpenTelemetry 1.45.2 for known vulnerabilities; ignore every .minibrain* dev-data folder (3b59859)
- data: a lone preview gets half the preview cores of one slot instead of one thread, and saves spill to a folder of their own instead of a shared one (3532e61)
- auth: turning an account off asks first and says how many API keys it deletes and that turning it back on does not restore them; accounts list their API key count (5faaac4)
- security: a storage test uses the saved S3 secret only for the saved endpoint and region, asks for the secret again otherwise, and is audited (88b7f73)
- security: the PKCS#11 library path is set only in minibrain.yaml or the environment; the API refuses to change it and ignores a stored value, and Settings shows it read-only (52ce9fa)
- data: all previews together stay within a quarter of memory and half the cores, split evenly over every run that can be going at once; a run that outgrows its share spills to a folder of its own, then says to run a smaller query or save (8444b21)
- compute: a pool with queued training or scoring jobs held to it cannot stop taking training; the change is refused with the waiting jobs named instead of stranding them (3d1bfab)
- worker: a worker deletes each job's log, outputs and attempt folder once main has them, and prunes work folders and downloaded inputs that no queued or running job needs (c0cf74f)
- data: personal-data fixes act on exactly what the check flags, and previews run per page with a server-wide limit (78f5949)
- worker: a worker keeps its job running and buffers its events while main is unreachable, waits for main to report the outcome, and keeps retrying instead of exiting (ac238ea)
- audit: secret-status markers such as kept and replaced stay readable in audit details instead of being redacted on write (c8ad2d1)
- data: dataset editor fixes: sample retry saves the current query, create sends details in one request with tag checks, the paused preview settles after 1.5 s and says it is waiting, abandoned preview runs are stopped, a fix applies once, Show rows ends when the query changes and counts the filtered rows, suggested columns no longer trip the leave guard, stale save errors clear on edit, SQL error marks survive surrounding whitespace, distinct counts read as estimates, and the dataset page lists its clean-up steps (07ef92c)
- api: worker claims and gateway config long-polls wait without holding a thread, so many idle workers no longer starve every other request (dce9812)
- security: the home folder and secrets are 0700, the database and its WAL files 0600, and a new secret is written once, atomically, with 0600 from the start (5b7b903)
- auth: turning an account off revokes the API keys it created, audited, and the gateway ignores keys whose creator is turned off (4a71402)
- audit: a break the chain page has not recorded yet points to Verify again, and other open breaks are not called earlier when the same check found them (48a589d)
- security: connection and encryption tests send a saved secret only to the saved address, refuse config-file connections, and are audited (1db406e)
- audit: checks walk past a break so later breaks are recorded, a row tampered with again after acknowledgement is raised again, the chain page records the breaks it sees, and secret values are redacted before they are stored and sealed (9c1a530)
- jobs: a cancel that loses the race to a claim asks the running job to stop instead of overwriting the lease, and two concurrent retries of the same work enqueue one job (fbdb6f0)
- compute: a draining machine keeps its engine until a machine that is not held has a ready replica, so draining every host never leaves a model unserved (f64b577)
- audit: a chain break not yet in the history says the next scheduled check records it, and each check that finds a break refreshes problems and the audit log (d9f54fa)
- infrastructure: CPU-only machines show no GPU memory, stay out of the GPU memory totals and sort below GPU machines (5e81a48)
- models: switching versions resets the version page and model card, so an open note draft never saves onto another version (2a9660d)
- playground: a save that finishes after New chat or after opening another chat no longer takes over the current conversation (4c13e60)
- auth: drop every cached query when the signed-in person changes, so the next person never sees the last one's name, role or sessions; profile forms reset per person (51e5551)
- gateway: request bodies over a per-path limit get 413 before they are parsed; source and artifact uploads stream unlimited, and gateways ship logs in parts (ee47813)
- auth: tokens and CSRF values with non-ASCII characters are refused instead of failing with a server error (914e445)
- serve: pool credentials get base weights and files only for models placed in their pool, and refusals are audited (4d6201a)
- auth: a trusted proxy's forwarded address reads every X-Forwarded-For line, so a spoofed first line no longer becomes the client (8723af2)
- auth: SSO matches an email to auth.admins and roles only when the provider marks it verified, not when email_verified is missing (b6fb5e0)
- data: the dataset editor waits while the completion list is open, shows SQL errors without DuckDB's excerpt or error kind and without an empty-rows line, keeps the red border for save errors, and drops the row count under Quality and Example (c470812)
- data: label balance puts each bar under its label, so label names show in full and bars span the panel (54c085f)
- playground: functions are added and edited in a dialog that validates the name and parameters, so a send no longer fails on them (a603221)
- data: the folder name dialog no longer flashes a name-taken error for the folder it just created (e4499ee)
- playground: only the conversation scrolls, never the page; New chat sits at the top of the Recent column when it shows (bee4856)
- connections: the copyable install command is a runnable command (0f0d8de)
- audit: Action sort knows the Full dataset version saved label (11766fb)
- connections: DuckDB extensions copy says minibrain never downloads them while running; install online or --from a folder (53b308e)
- monitoring: Requests indicator reads Paused while sorted instead of Live off the newest order (4983be2)
- monitoring: Requests pages keep the time window of the first page, so totals hold (e73ffb7)
- lists: Jobs Kind and Audit log Action sort by the label shown, unknown codes last (a427103)
- exports: retrying an old export reuses a ready one and refuses a changed base (bfc304a)
- versions: deploy refuses a version whose base is no longer on the install (303756e)
- models: the list says a version is ready by the same rule as the model page (850494e)
- overview: the usage strip always shows four tiles; cost reads – without traffic (4d26c11)
- home: Jobs running counts running jobs across the install, not queued ones on custom models (31e0867)
- home: Serving counts deployed models, not every model that once had a live version (21614f8)
- job-page: Result keeps its size while loading, with placeholder scores, instead of an empty box that grows (5cbd557)
- monitoring: Requests hides Finish and Response ID by default, like the model page, so the grid fits 1280 (4cc12bf)
- api-keys: Name truncates at 192px so the grid fits 1280 with generated key names (311066e)
- data: label balance over 200 labels says it shows the top 200 and gives shares of every labelled row (df697ac)
- data: empty text cells sort last in previews and SQL results, like missing ones (691352b)
- data: deleting a folder with one item says it moves up, not move up (6fdd3b8)
- job-page: Result's verdict line truncates inside its column instead of running under the scores (d1afaae)
- job-page: Ran on names the pool and machine the latest attempt ran on, not where its host is now (64e29f5)
- data: a query that joins the source alias with other sources is refused alike by Run and Save, naming the tables to use (1fd9022)
- machines: an approved machine offline for a week stays listed until removed, so its approval can still be revoked (37cd547)
- monitoring: the Cost / 1M tokens axis reads in GPU hours or seconds like its headline instead of bare GPU-second counts (f2d822f)
- machines: a pool token writes join_token.joined once per machine, not on every call when machines sharing it alternate (fd5ab3d)
- extensions: minibrain extensions install with no source installs from DuckDB's core repository, which needs no httpfs to fetch httpfs (a85bf37)
- data: a dataset version with too few rows to split says so instead of blaming a change that never happened (a4ce9e5)
- jobs: the Pool column and its sort name the pool each job ran in, not its worker's current pool (c63f097)
- infrastructure: This main process no longer claims training avoids the Apple GPU; it trains there (9608c87)
- data: dataset preview counts empty rows, duplicate inputs and labels after stripping whitespace, like checks and training (1321404)
- jobs: sorting by Duration pages without duplicates or gaps while jobs run; the cursor keeps the moment the first page measured from (d4b490b)
- model-overview: the next-step line offers only a version that meets the deploy rule, so its Deploy never opens on a blocked version (94ce461)
- model-card: Ready to deploy on the card is the deploy rule and not live, like the Overview: an unscored ready version that meets the rule reads Ready to deploy, a version on another base does not (43f13a0)
- data: Date columns sort in previews and SQL results; a date's day no longer reads as a time zone offset (c504113)
- machines: long host and pool names in Waiting for approval truncate, so Approve and Deny stay on screen (71a1606)
- settings: default training recipe labels in sentence case (f24368e)
- api-keys: key ID gets its own column; Name narrows; the grid fits 1280 (3fa73c4)
- playground: Recent, chat and settings fill the screen height and end at one bottom edge (e34e335)
- home: wide page frame like other pages; Get started beside Needs attention; overview counts move into the header (a4e7775)
- feedback: text columns take the width; Status is capped at 264px (7aba5bd)
- monitoring: model filter and time range share one control bar; stats fill an even grid; no traffic is one line inside the stats card (0bb6a74)
- model-overview: About gives the live version note and description full-width rows; Current setup in two columns with Integrity full width (799a196)
- jobs: Progress and Inputs end at the same height; Result puts the verdict and actions beside its scores (7793914)
- infrastructure: Setup puts This main process beside Loaded in memory and caps the diagram's connectors; Needs attention sizes to its rows; Running cells show all activity in a tooltip (7078da0)
- models: Ready to deploy means the version meets the deploy rule and is not live, the same on the card, Overview and Versions tab (e2708f5)
- model-overview: the next-step line offers the best ready version and only suggests training while v0 is the only one (61950c2)
- models: Deploy's disabled hint says the newest version is training or being scored instead of no version is ready (860623b)
- auth: workers download only what the jobs they hold need, engines only adapters served in their pool (3fa2f9c)
- auth: engine tokens are bound to the join token that admitted their worker, so revoking or expiring it ends them (081dcf1)
- ui: in-progress status dots ping, fill or wait: fleet jobs and loading engines, draining machines, key rotation, audit chain check, deploying, deleting and the version verdict line (17e43a8)
- ui: rows whose status dot carries the progress arc drop the duplicate progress bar (models list, Versions tab, Jobs, Discover, teardowns) (2db7f9a)
- models: a version whose training or scoring job waits for a worker shows Queued with the dashed ring in the models list and Versions tab (6cd896f)
- models: View options lets a column hidden for lack of room be switched off, keeping the Hidden: not enough room note (1e472f8)
- api: ids beyond the signed 64-bit range in any numeric path parameter return 404 instead of 500 (a6d134d)
- infrastructure: machine Live sparklines use fixed axes, GPU busy 0-100% and memory 0-capacity, with now and the hour's peak beside them (9190134)
- auth: OIDC group claims get the proxy's caps: 200 groups of at most 200 characters (282415f)
- web: the shell does not poll models for people without a role, so Your account makes no 403 calls (4600f72)
- compute: running and recent jobs belong to the machine each attempt ran on, not the worker's current pool (eaa21a8)
- infrastructure: a machine's problem badge takes the worst problem's tone, so a failed engine shows red (9103a5b)
- models: Repair it and the Infrastructure link in Where to run it open a new tab on Add model, keeping the form (53383cd)
- jobs: lease expiry fails a running job that has no attempt row instead of stopping expiry for every job (d9b9f7e)
- models: View options marks columns the table had to drop as Hidden: not enough room (7858d8f)
- data: text files with a UTF-16 byte order mark are read as UTF-16 instead of garbled Windows-1252 (95ae6fc)
- infrastructure: rejoin commands and the run-elsewhere help name the pool's join token in MINIBRAIN_JOIN_TOKEN (388a95b)
- data: an upload that is only an empty JSON array is refused like an empty file (df601c4)
- discover: Compare and a model's Formats table fit at 1024 wide: engines join the format cell, memory cells wrap (8775671)
- models: Usage panel footer names the pool only, without the machine hostname (584bbea)
- models: bulk deploy tries every variant, refreshes the list and shows what deployed and what failed with reasons (8f00530)
- versions: Cancel shows Cancelling… at once, and a run that finished before the cancel landed says so instead of Cancelled (348dab9)
- versions: a failed version's page offers Retry next to its job, retraining or rescoring like the Versions table (6c6ea5c)
- security: policy descriptions are one short line; the detail moves into an info tip beside each label (7821174)
- models: a deployed model without data shows Deployed with a Needs data hint instead of hiding its live state (d27f80e)
- training: what changed compares a run against its preset (Balanced for custom), not the install defaults, and skips a row limit that kept every row (11bc72c)
- settings: Storage's Save and Test connection sit in the card footer like every other settings section (c45a501)
- models: Add model's dataset menu counts a sampled version's rows, not the whole source (a939c60)
- models: demo data over the row limit on Add model offers a seeded random sample instead of only an error (e2de287)
- data: taking a sample inside the Add model form no longer submits the form, and Enter in the rows box takes the sample (fe3da11)
- web: an unknown address shows Page not found with its section's crumb, not the nearest page's title (30f4c13)
- web: in-page not-found states use the page header layout, so their title sits where every page title sits (b5e4027)
- models: the Compare pickers name an unscored or repeated version instead of showing a bare number (9088f83)
- models: Add model on a dataset with no version over the row limit offers a seeded random sample (540724b)
- audit: session counts in account events read 1 other session, not 1 other sessions (21a9c42)
- account: a refused password change is audited with its reason, like a refused sign-in (b0f38b2)
- data: a Windows-1252 file that starts with a UTF-8 mark no longer gets it in its first column name (43ac835)
- models: the version page disables Delete version on v0, the baseline the API refuses to delete (c2cd602)
- jobs: a failed or cancelled run shows no score, not the score its retry later earned (a685f31)
- models: a version's Rows trained is the count its run measured, feedback rows included (813ee63)
- training: the Train page says a run is scored against live or v0, as its version page compares it (8c48ccc)
- accounts: names drop control and text-direction characters so they cannot hide or reorder text in the audit log (9d5ea09)
- data: an encrypted file read on a fresh DuckDB thread continues from the last position, so uploading [] no longer spins the server (0443a44)
- account: searching for a profile setting lands on and highlights the Profile section (5278d66)
- account: session lists, counts and sign-out totals skip expired and idle sessions (3706ce8)
- web: a link to a pool whose name has a space or other escaped character scrolls to that pool (f5a209f)
- infrastructure: a low disk is one issue per disk, named by its mount, and Storage warns at the same threshold as Needs a look (77b2e73)
- infrastructure: the rejoin hint for an engine-only machine names the gateway token it needs, not the worker token (5ebd815)
- web: a pool's hover card opens Infrastructure in a new tab on create pages, so the form keeps its state (a2264d0)
- infrastructure: Free up space no longer offers a base model that is a judge or safety model (2d05e89)
- data: the sample offer names what is over the limit: this file, this source or this dataset, not always this query (07cebd0)
- models: a version row in the models list opens that version's page (73ab3cb)
- training: the Training picker is as wide as the other fields and its description truncates instead of spilling out (ca3303b)
- models: a cancelled run shows as Cancelled, not Failed with no reason recorded (82a8dae)
- jobs: Back from a job page returns to a version page instead of bouncing to the job (da08c94)
- models: the version page shows one Loss and one Learning rate chart while training, not a new pair per refresh (ff90094)
- models: model cards and AI-BOMs keep exact scores and round only for display, so they match the model page (ae528e0)
- models: activity names and groups people by account id, so two accounts sharing an email stay apart (4ab3fcb)
- training: the run summary counts the accepted feedback rows the run adds (3dd438a)
- models: the Model column keeps at least 15rem so names and version labels stop truncating at 1280 and overlapping Data (1183248)
- auth: sign-in attempts are throttled per client address and lockouts hold only for the address that guessed (8506ed0)
- models: scores round half up to one decimal the same way in the UI, status lines, model cards and AI-BOMs (ca50786)
- models: Discover rows and model pages show Add model disabled with the offline reason when offline (8d4839d)
- models: a bulk delete that fails partway refreshes the list and shows what was deleted and what failed (05ee38d)
- models: a base model's page shows its download or import progress instead of Not deployed (beccd5f)
- models: activity says v0 is queued for scoring until its scoring finishes, then that it was scored (ff124bf)
- models: activity names a deleted version as deleted instead of linking to a missing page (3589963)
- models: every failed version can be retried, retraining or rescoring as its failure needs (11e13a1)
- jobs: the event stream reads the job status before its events so the final status line is never missed (57b3b8e)
- jobs: lease expiry locks the job rows it reads so it cannot overwrite a concurrent heartbeat (d2a6893)
- jobs: a retried job's time limit counts from the start of its current attempt (152b6de)
- data: the upload error lists JSON, file sizes under 1 KB are exact, and the output suggestion says when no answer column is picked (0f256e8)
- models: the Train page says how many accepted feedback rows the run adds, counted the way the run merges them (9995dda)
- data: an upload matches an earlier one only when both content and file type are the same (40d5c81)
- data: empty files are refused at upload, and a source with no rows is refused as a dataset with one message (6cd6e86)
- data: rows up to 64 MB are read, and a longer row gets a plain error instead of DuckDB's (5d971ce)
- data: text files that are not valid UTF-8 are read as Windows-1252, and the source page names the encoding (f1be7b8)
- accounts: a long unbroken name is cut short instead of pushing the other columns off the table (7f0c1df)
- auth: the sign-in page shows only fixed single sign-on messages, never text from the URL (99fbed2)
- jobs: a job's event stream stops once its reader is signed out or turned off (c4f8351)
- auth: API keys and audit events belong to the account id, so a display name can no longer pose as someone else (1f2d3b6)
- models: a key created from the Connect dialog updates the models list key count (e59b85d)
- keys: the key dialog offers only expiries within the install's maximum key lifetime (4651fbb)
- models: bulk deploy skips variants that can't run here and shows why (b44ed7f)
- models: Deploy vN on the models list opens the full Deploy dialog with pool and replicas (b870eaf)
- gateway: a model being undeployed stops answering on its versions, so no client reloads an adapter and stalls the stop (d2e9db3)
- data: deleting a dataset is refused while a training or scoring job on it is queued or running (5107e3e)
- gateway: an engine that reports ready but cannot be reached fails in 30s with a 503 naming its pool (d03d00f)
- models: Train vN names the number the next version really gets after a version is deleted (cc0da41)
- audit: the audit log names base model downloads and imports (11d6cb1)
- models: starting a base model download or import is audited, so its page names who added it (16baf5a)
- models: the quality tooltip calls a version best only when it beats the one shown (4addad6)
- models: a model page whose versions fail to load says so instead of showing no versions (b08aa85)
- policy: a cleared text or host policy is blank instead of the word None, and lists are refused (dfd4447)
- data: replacing an uploaded file refreshes its rows, column stats and dataset previews (6e7adf5)
- data: drop repeated inputs that differ only in surrounding spaces, as the repeated-inputs check promises (6ca61a6)
- data: keep label case in check details instead of lowercasing the whole sentence (be1addb)
- data: read TSV files with a tab delimiter so text with commas keeps its columns (324af0e)
- auth: sign-in takes as long for an unknown email as for a wrong password, so timing does not reveal accounts (7b74d6d)
- models: deleting versions with v0 ticked no longer says a deployed vnull is kept (6fa9417)
- models: renaming a model needs a letter or number, as creating it does (a938321)
- jobs: stack a job's progress and settings panels below 1280 px so times stop wrapping letter by letter (10884c5)
- web: a one-day audit retention reads 1 day, not 1 days (d8e7fd2)
- gateway: a stream the caller closes is logged as 499, not as a 500 server error (9bde74b)
- versions: say why a version's retry was refused instead of failing silently (18194d0)
- app: send nosniff, no-framing and same-origin referrer headers on every response (a2a1053)
- models: model cards name the metric they score by and no longer report a macro F1 of 0 for free-text models (54e9bb9)
- jobs: fail a job whose result cannot be saved instead of leaving it running until its lease expires and it reruns (3a7dffa)
- auth: a short idle timeout no longer signs out people who are using minibrain (3ea19a2)
- deploy: deploying on the same pool keeps its replicas while a machine is offline instead of scaling down (0ea737f)
- app: a crash answers 500 with its request id, and minibrain warnings reach the console again (0879f0e)
- engines: an engine token downloads adapters only, not datasets or logs (8ba7e16)
- evaluation: a rescore with changed graders replaces the one still running instead of reusing its old graders (65fda87)
- web: unknown Infrastructure sections show not found instead of the overview (39841a7)
- playground: View code points to API keys, not a Deployments page that does not exist (769bcff)
- gateway: calling an undeployed model says it is not deployed and names a version to call (2f7ad40)
- models: a version rolled back from is not suggested when it scores below live (e74728a)
- monitoring: replicas list only the live version's base in the model's pools (0a2d7bb)
- models: models list never offers to delete v0, which the server refuses (87a1c90)
- auth: ignore an OIDC email the provider marks unverified so it cannot match auth.admins (b0ce119)
- worker: keep a job's lease while it fetches inputs and uploads the adapter, so slow transfers are not lost and rerun (0698d4d)
- policy: refuse policy values of the wrong type instead of failing with a 500 (0ad5b56)
- jobs: retry a job once and never beside an active run of the same work, so a version is not built twice or retrained while live (82881f1)
- app: close the database pool on shutdown and keep the key cache from pinning an app's engine (a586cca)
- ui: page titles truncate on one line so header actions stay on the right; captions run under the actions (78ba741)
- ui: not-found, crash and no-access pages get the page header and crumb row; Add model gets a caption (350500f)
- models: version crumbs never show vNaN; a missing or unparsable version reads Version, compare without a pair picks live and newest (00033a8)
- models: list fits its width: tighter tracks, low-priority columns hide when there is no room, idle bars clip instead of wrapping (0a6863d)
- playground: drop the Recent panel's duplicate New chat button and the Safety model setting (633d974)
- playground: chats save by default; only an install-wide lock on request content stops them (85ae0b0)
- api: 404 details for models, versions, jobs and requests read as sentences (2e7b816)
- keys: API key scheme names the API keys page (1d5d063)
- models: Activity shows only this model's events; status line never offers the v0 baseline (c8c4142)
- web: "fits your compute" and "too large to train" for the capacity check (8bfa40c)
- web: "runs on your hardware" instead of "runs here", since the check covers every compute pool (8db1e6d)
- proto: style tab previews keep their full size inside the kit's tabs (7328fc6)
- web: job links on the models list and delete dialog open training runs or job pages by kind (0fcf818)
- web: model page on main's job links, training links and breadcrumbs (2412e95)
- models: the way out of a model that can't run here is View details (7264fda)
- models: a model that can't run here offers to add the named variant that can, opening Add model with it picked (103b4ff)
- web: clear errors when the server is down or fails, and no upload to a removed route (1c70b4b)
- discover: the table fits again; formats move under the model name and run and train memory share one column (5d81b67)
- models: the base picker shows a base being deleted as Deleting and won't pick it (c9caad2)
- models: a base being deleted takes no new models, and a failed delete starts again with the current ones (be5b25d)
- gateway: a main that stops cleanly removes its own gateway report (a70f033)
- cli: Ctrl+C stops minibrain and mise run dev with one line, no tracebacks (6bde948)
- models: Add model offers only Classify and lists the other tasks under Coming; the roadmap no longer claims a chat preview or a teacher eval (517e319)
- web: register the prototype pages only in development, so production builds leave them out (6241c60)
- data: an S3 or Iceberg path with or without a trailing slash is the same source (d390d27)
- engines: fetch a base's working copy as its own step with progress, so the startup window starts when loading does (f37c59b)
- models: refuse training on a quantized base in the bucket up front, reading its config.json through the store (8693c43)
- theme: clear Cobalt properties the new mode does not set, so an OS switch to light leaves no dark colours (c2c4ba8)
- worker: job format 2, so older workers cannot take jobs that stage files for the bucket (47ad6c5)
- storage: the folder and the bucket follow one write-once contract (cb45910)
- models: the base picker scrolls only inside its list, and the Import tab inside the dialog (ed355f4)
- models: quantized bases, FP8 included, are refused for training up front (75b83cf)
- data: deleting sources that datasets read offers to delete those datasets too (79f1e08)
- data: add the dataset versions and version diff components the last commit left out (d146981)
- models: a base model counts only requests made since it was added (a356041)
- models: bulk delete removes each version once and treats an already deleted version as done (f5d64c1)
- models: Add model buttons sit at the end of the form instead of sticking to the bottom (10d2031)
- discover: the loading table fills the page (ae61acf)
- discover: loading skeleton matches the loaded filters and table; Models row menu says Edit (9757b61)
Performance
- encryption: re-wrap a bucket object's key header with a server-side multipart copy, sending 5 MiB instead of the whole object (5b64f5e)
Refactoring
- models: drop the next-step notice from the model overview (7762701)
- web: Tabler icons replace lucide-react; obp-ui updated to the Tabler kit (0dd41c4)
- web: delete the unused TrainingData page and the dataset mapping and query hook only it used (19a9fff)
- ui: one UsageGauge for memory and disk bars on Infrastructure, Add model and Train (1bc2498)
- data: sample data is demo data everywhere: /api/demo-data, POST /api/models/demo, the demo_data folder and Add › Demo data (1fac33a)
- models: model, version, compare and train pages use the shared page header and group-rooted crumbs (c4bb8dd)
- training: remove the Train dialog now that training is a page (9a64c04)
- web: call the merged models API (fdab154)
- api: one models resource keyed by the API name (eaf1982)
- proto: /proto/model-icons keeps four framed avatar styles (0c7e959)
- web: the test examples query takes only the model (a49ab94)
- take recipes out; Try with sample data makes a Classify model on the travel intents sample (9484283)
- catalog: the default base's description says models, not minibrains (3de5bfd)
- snapshot becomes dataset version in code, API and docs (4cac648)
- Minibrain becomes Model and the old Model becomes BaseModel (62662e4)
- db: squash the migrations into one baseline (c98d4e4)
- delete every path kept for old installs (b84e9b4)
- delete Copy to the bucket, the job format check, inline OIDC and collector secrets, and three routes no UI used (1c4b907)
- models: base models stay until deleted; drop the cache budget, pins and remove-files (d9f4f42)
Documentation
- roadmap: v1 light-mode pass; log the flaky long-poll timing test under Fixes (00a38d0)
- rewrite the README around the quick start, the journey and the endpoint, with status badges; minibrain installs from PyPI (99c7eb7)
- api: the preview endpoint describes the per-slot core split (9ba5609)
- roadmap: before-release ops items: deploy manifests, alert rules with runbooks, rolling upgrades, cross-component trace ids (e1085a7)
- product: dataset checks live under Data quality (738b5a0)
- roadmap: the before-release list for v1: page reviews, security, docs, logo, website and release checks (903184d)
- roadmap: teacher labelling, generated Q&A, Distill, the teacher pool and the private frontier endpoint move to v2; v1 trains on data that already has answers (9d0340e)
- plan a global ⌘K search for v2 alongside the top bar and notifications (ab1cfd3)
- security: the Python SBOM extras list names every extra CI exports (8b1fd3d)
- roadmap: the core journey every UI change serves (3d8036c)
- roadmap: dataset required at creation, v0 untrained baseline, model Settings tab with re-scoring, Train dialog from the last run (df14579)
- roadmap: a custom model is bound to one dataset; changing it is a later item (9ce60b6)
- roadmap: the platform as six things and three rules; deploy and undeploy run as jobs (28068e0)
- roadmap: embeddings move to retrieval as their own model kind (a6400ef)
- the model overview no longer shows a next step (aa401d0)
- roadmap: managed compute in v3, opt-in; self-hosted and air-gapped stay the default (753d2e7)
- roadmap: compose and a Helm chart ship in v1; the Kubernetes launcher stays in v2 (689a16c)
- roadmap: the general instruction mix the training section promises is not implemented yet (72b6f3f)
- roadmap: pools say what they run (serving, jobs or both), with serving first on shared pools (027b393)
- drop the icon exceptions now that the portfolio has no icon rule (67b2ef1)
- roadmap: a large datasets milestone for streaming, multi-GPU, full fine-tuning and continued pretraining (05db7ce)
- roadmap: continue training from a version, and make a custom model a base by merging it (fb18eca)
- roadmap: product priorities, a drop-in private endpoint and retrieval first (b894403)
- serving: machines report Intel XPU as an accelerator (f38695d)
- roadmap: queue the Minibrain to Model rename for after the task engine (70c13b1)
- roadmap: MLX training on Macs parked for later, behind a flag (4b0bd83)
- roadmap: park the larger ideas under v2 until the task engine settles (7ce780c)
- product: docs/product.md, an internal inventory of every feature and use case (41ea37d)
- roadmap: a Fixes list for known defects and leftovers to delete (44cb632)
- storage: base weights in the bucket, signed links, data keys and working copies (2787659)
- config: minibrain.yaml reference with a full annotated example (a2806f9)
- roadmap: recipes, worked use cases with a config, sample data and a guide, started from Add model (6586aea)
- roadmap: build the task engine first as Custom, then presets, with no compatibility layer (f8cbb5f)
- roadmap: task types are presets over an output contract and graders (5adecaa)
- roadmap: base models stay until deleted; the cache budget, pins and remove-files are gone (9f2b3b0)
- roadmap: logo rule for base and custom models; custom logos in v2 (e491313)
- Cobalt's type pairing is Mona Sans with JetBrains Mono (9717bae)
Build
- web: obp-ui with tooltips spaced off their trigger (a06d2b4)
- web: obp-ui with tooltips on a collapsed sidebar (70d9664)
- web: obp-ui with a composer focus that never scrolls the page (c15c06f)
- web: update obp-ui for dark tooltips and InfoTips (881baa1)
- web: take obp-ui with larger theme buttons (f70558c)
- web: take obp-ui with square, larger tabs (45d8375)
- web: take obp-ui with text-base as the default size (3f73e65)
- web: take obp-ui with the LogViewer (1e24b82)
- deps: transformers 5.22.1 and TRL 1.17.1 (6d5971a)
- web: obp-ui with the new toast (a38bb80)
- web: obp-ui with underlines clear of descenders (982567c)
- web: obp-ui with a centred empty state in tabbed search lists (05e9842)
CI
- tests run in parallel from a pre-migrated template database; a release builds images while CI runs and skips the tests a commit already passed (a559aca)
- release: publish the wheel and sdist to PyPI through Trusted Publishing after the GitHub Release (b83ceb8)
- check Python and web dependency licences against the Apache-2.0 policy (b9967c6)
- product: check_product.py fails when a route, command, config section or policy is missing from docs/product.md (e2bfb46)
Tests
- access: list the key expiry route as builder-only (206d8dd)
- access: call every guarded /api route as each lower role and every /v1 and internal route without its token (5692a88)
- gateway: the Connect dialog's test request works with a key scoped to the model, and 401/403/404 carry the messages it shows (b924467)
- access: list the rescore route's role (93bb9aa)
- encryption: narrow the bucket to BucketBlobs in the header re-wrap test, so pyright passes (5cd4b89)
- storage: run the storage tests on the folder and on a moto S3 bucket (d9e73e3)
Style
- tests: format the release-notes test after the organisation rename (a91849a)
- data: the SQL editor's tables box reads as Sources, with each source's icon and row count (89ab161)
- web: every header button carries an icon (058c446)
- infrastructure: Delete logs is a destructive button with a trash icon (ea4e716)
- infrastructure: machines table cells share their header's edges; pool bands sit on the column grid without room hints or totals, sort headers use the shared SortButton, and the page drops Needs attention (c135d22)
- data: About reads Overview, Data quality, Query, Label balance, Used by; label shares show count and percent, dismiss and edit query are icon buttons, Used by links the newest model version, train and version buttons carry icons (bf99e8a)
- models: deploy dialogs use a compact label column and open pool links in a new tab (ea610f2)
- playground: widen the settings column to 25rem so the preset buttons sit on one row (0dd67ac)
- data: Data quality, Label balance and Used by share one list style: 32px rows, no boxes, trailing values aligned right (c420c8b)
- data: About facts read label: value like the query, each split on its own row, and section content sits indented under its title (c62bf16)
- data: Add goes back to a + icon button with its label in the tooltip (2e1c586)
- data: the catalogue search reads Search… and Add is a labelled button (6cfffb7)
- data: the dataset's checks are titled Data quality and the separate quality toggle goes; info icons are larger (bf7af93)
- data: About sections are set apart by space, not dividers (8f4cd94)
- data: the S3 source group is Object storage (S3, MinIO, Ceph); its action is Add bucket path (609b290)
- data: About sections are divided and evenly spaced; the dataset Overview leads with rows, labels and sources, splits on one line and data quality behind a toggle (2851aec)
- data: Used by lists models with their avatars; the test examples note is a short +N with the detail in a tooltip (ed81ff5)
- shell: the account tile uses the kit's muted tone instead of the brand colour (df5f584)
- models: the Models list status says only what runs (Deployed, Not deployed, training, failed); readiness to deploy lives on the model's page (f8580ce)
- data: N passed on dataset checks is a visible tooltip trigger listing each passed check with its detail (36868b5)
- data: Used by on a dataset is a read-only list of the models that train on it; the dropdown that re-pointed other models goes (e6b9525)
- versions: every row's actions sit in one column at the end: Deploy on ready rows, Retry on failed ones, the … menu on all, with Deploy for other versions inside it (bfe25fe)
- versions: format the client import (f5be83c)
- versions: drop the Scored by line above the Versions table; scoring setup lives in Evaluation (5284da5)
- infrastructure: pools, machines and models carry their icons in the Machines table; memory shows the share with measured or estimated in a tooltip (41af370)
- data: an empty folder points at its New dataset button (9c06f8f)
- ui: a ready version that is not live reads Not deployed instead of Awaiting promote (3b40233)
- add-model: Pools becomes Memory with a pool picker that starts on the best fit; Disk becomes Storage (1d09296)
- ui: disk and storage icons are a cylinder instead of an SD card (a01d156)
- version-page: drop Copy from the … menu; Summary's API name has the copy button (ddaeb3f)
- version-page: the API name's copy button says Copy API name (f293322)
- version-page: the deploy rule and the version note move into Summary as facts; the verdict and note lines above it go (8276288)
- ui: pages end with 48px of space below their last section (02bdfdb)
- settings: the links to Security on This install carry the Security icon (fff6d6a)
- settings: This install drops Copy versions; Download diagnostics and Open Infrastructure carry their icons (982c6f5)
- settings: adding or editing a connection opens a dialog; the connections list keeps a fixed height and scrolls (3523969)
- data: drop the source-dataset-models trail from item headers; each page's Used by section already links them (29408f2)
- data: a source's header line shows only when it was added; the tree already places it and the SQL editor lists table names (c811e9c)
- ui: page headers show their icon in a 48px tile beside title and caption, the same shape as a model's avatar (2afcaf5)
- training: version sits under the model name and Duration is hidden by default, so the Training list fits 1280px; disk icons read as storage, not servers (fc1719e)
- account: shorter section descriptions on Your account (30d13d2)
- models: scores read vs baseline and vs live; the deploy prompt says better than live (3e57073)
- training: Train a model carries the training icon; the Method column and filter stay hidden while SFT is the only method (1cff6b2)
- data: the Data header spans the full width like the catalogue and detail below it (c9fecb7)
- ui: memory and disk bars run the full width with the used share above their right end (a8235b1)
- ui: memory and disk bars are 20px with joined parts, hairline seams, an outlined planned part and the used share beside them (d32b596)
- training: the compared run's training and held-out loss get their own muted hues and dashed lines (7dcc20a)
- models: model settings sit in a 768px column so rows and descriptions stay a readable length (b0b679b)
- jobs: Cancel on a job is an outlined destructive button (obp-ui destructive-outline) (c81b74b)
- models: the version run panel keeps one link to the job, dropping the duplicate Logs button (ee192ee)
- web: the summary panel on create pages is as tall as its content, not the form (014e735)
- models: the models search sits next to Add model in the page header (15ee051)
- web: page header actions sit on the bottom line of the title and caption (d9ee42f)
- web: 32px between the crumbs and the page title (bf7d318)
- web: more room between the crumbs and the page title (df1588f)
- web: format the avatar tile (2bcc5bd)
- models: avatars drop their border, every style sits on a tint of its colour, and publisher logos stand without a tile (b3c9f34)
- playground: plainer note when chats are not saved, pointing at the model's Settings (0596ef5)
- web: format ApiKeys actions (a37f72a)
- models: drop the All API keys link from a model's API keys tab (bd3233e)
- models: leaner API Connect dialog; feedback call is a code tab (38aa436)
- models: Monitoring range sits in the Traffic chart; one empty message that says what to do next (635e40f)
- models: plainer serve-only notice on Add model (ddf0427)
- models: shorter empty Models message; Try with demo data (0edb23d)
- models: drop the Download row from system requirements; it repeated Storage (7cc3ccf)
- models: licence acceptance is the last row on Add model (f0eb91a)
- models: memory and disk icons before the system requirement amounts on Add model (3710ee9)
- web: roomier error panel with the status code and path in a footer (5564809)
- web: format Home (3d579ba)
- models: view options in the table header, and the table header stays when there are no models (da8076c)
- models: no Name kept note; the name shuffle always replaces the name (ad5c633)
- models: API name in a disabled input and a 64px avatar on Add model (cf87c51)
- models: Add model gets an API name row and a clickable avatar, and the name shuffle no longer changes the avatar (bb8dfd3)
- models: Add model shows the avatar large in its own row, and create-page footers use base-size text (bffbd4c)
- models: create-page footers only speak up for problems; the summary card holds the result (ebe4efc)
- models: status icons by the model name, a large avatar in Settings, API name only in the header, and Duplicate and Switch base out of the Danger zone (5f7d60a)
- models: base text size in the create pages' summary card (bfd138d)
- training: format the pool summary (3b0a4da)
- training: no likely-machine line under the pool (ead92a0)
- training: the compute row is Pool, matching Infrastructure (90b34d6)
- models: Result instead of What you'll get on create summaries (0509891)
- models: breadcrumbs on every model page header, and Show parameters on the Training card (6e5a50f)
- web: an icon on the create pages' Cancel button (bbaab19)
- training: two-line preset dropdown; docs(roadmap): training from code via API, CLI and rules (ec9ad6c)
- training: larger preset dropdown with larger icons, and the icon shown when selected (c998d7a)
- web: form spacing defined once in FormRows and create-page actions span the page (0f3c504)
- models: view options at the right end of the toolbar (fff0611)
- training: training preset as a left dropdown, narrower parameters pane, clearer title (87244d3)
- proto: Use menu without descriptions, API Connect, one Change deployment item (1ed8e7a)
- models: train dialog without the from-base line or coming-soon methods (6068533)
- proto: Use is an outline button like Train and Deploy (5ed5b02)
- keys: the API keys page lists keys without a section heading (561d9a7)
- models: tidy the versions actions (a35ed8b)
- models: version actions appear only while versions are selected (e45e3c8)
- models: versions buttons carry icons, and delete is destructive (bb129a4)
- models: add model side panel with format, system requirements and memory and storage on your compute (7a9d67c)
- models: no Custom badge or Not deployed label in the model header (77ca42e)
- models: no naming hint under the model name (0f9b743)
- models: avatar picker explains the benefit, keeps one height, and drops the avatar undo note (129e629)
- models: no base caption on the Versions section (ce99d29)
- models: usage controls beside the section title, no hint line, and no total while there is no traffic (575bb0a)
- web: Settings only in the account menu, not the sidebar (1b4bd1f)
- models: base model chevron in the primary colour (8c28bb5)
- models: Add model has no caption (b0fcba2)
- models: shorter Add model caption (962d315)
- models: base model field shows a larger chevron and no Change label (2d32ae0)
- models: base picker details as a "Details" text link that opens in a new tab (38041f3)
- models: base picker rows show details as a hover info icon by the name, with no reserved space (fdee38c)
- proto: avatar picker without compare, twelve initials and larger style tabs (591028d)
- proto: no keyboard hint line in the avatar picker (d120053)
- proto: quieter group headings on the models list, title and count only (108be83)
- proto: overview as Usage, About and Details sections with outside titles; details always shown (795ccdf)
- web: more room above and below section titles (d4cb7f5)
- proto: page sections take the app's outside title, icon and description; overview widgets keep theirs inside (255090f)
- proto: more room around the model and version header, with its actions centred (b078dbc)
- web: text-base for body, rows, labels and values; text-sm only for hints, captions and meta (7f8649d)
- web: row buttons on model versions, Evaluations and Deployments are default size (827dc10)
- models: row actions on the Models page are default-size buttons (d94605c)
- models: row actions on the Models page are secondary buttons with an icon each (7b4632d)
- models: the picker's Details control is a labelled button, not a bare icon (a4effbd)
- models: picker tabs are 14px in a 40px row with their counts; drop the tabs prototype (023dde1)
- models: every Add model button uses the CubePlus icon (20bb0ca)
- models: base picker shows memory and size as icon and number, and the licence by name (5429c18)
- models: one SHA-256 column on a model's files, with a tick when it matches (5ddc2fa)
- models: group titles in My Models are quiet labels, not a second header (af592e1)
- models: requests show seven daily bars beside the total (ffae304)
- models: a custom model always shows a letter tile, a base its publisher logo (c685b2c)
- discover: logos sit centred in the model cell (1dc505d)
- models: larger publisher logos in Discover rows and model page headers (70faf60)
- discover: row actions are default-size buttons and variant Add model is secondary (8045c84)
- discover: Add model in the variants table is a default-size button (deecb1f)
- models: empty Models list is an icon, one line and Try the sample (85eeebe)
- discover: filters panel takes the 24px panel padding on a wider sidebar (a4b8428)
- models: empty Models list and every empty grid are centred in a taller area (4a8ed92)
- discover: roomier filters panel, and its skeleton to match (2d68cda)
- models: Deleting shows in red with a red progress bar (f6d00dd)
- type: page titles at normal weight; obp-ui with default font smoothing and roomier labels (545abc9)
- models: Add model fields share one width (17d9f32)
- models: Add model form and side panel always end at the same line (9c4034d)
- models: Add model disk and memory meters are tall continuous bars with a percentage and labels (d6cec3e)
- models: disk and memory as taller segmented meters on Add model (0096e1b)
Chores
- release: v0.1.1 (397466e)
- move the repository to the getminibrain organisation (87a76d7)
- web: delete the prototype pages and the code only they used (36f919f)
- samples: prefix the demo datasets with demo_ (ef9a735)
- keep agent instructions out of the repo (cbe9024)
- web: remove /proto/type; Cobalt keeps the Mona pairing (ddfbab0)
- dev: mise run setup-link prints a fresh first-admin link for .minibrain/ (68c1391)
Other changes
- Revert "style(discover): filters panel takes the 24px panel padding on a wider sidebar" (26514a0)
Full changelog: v0.1.0...v0.1.1