Releases: nayrosk/overbrainer
Releases · nayrosk/overbrainer
Release list
v0.4.1
Immutable
release. Only release title and notes can be modified.
v0.4.0
Immutable
release. Only release title and notes can be modified.
Added
- (cli) add overbrainer migrate for projects made before 0.4.0 (#71)
- (tui) add an init wizard and an auto mode (#67)
- (metrics) serve Prometheus metrics (#66)
- (config) reload overbrainer.toml and .env while running (#65)
- (runpod) choose GPUs, data centers, volumes and templates from the catalog (#60)
- (tui) add a Project view that shows and edits the configuration (#58)
- tell the user when a newer overbrainer is on crates.io (#53)
- (tui) show the version, GitHub link and live cost, merge the question detail, export logs (#47)
- record stage history and allow one overbrainer per project (#44)
Fixed
v0.3.1
Immutable
release. Only release title and notes can be modified.
v0.3.0
Immutable
release. Only release title and notes can be modified.
Fixed
- (tui) preserve stage data, fix resume counts, parse more list shapes, populate logs (#33)
v0.2.1
Immutable
release. Only release title and notes can be modified.
v0.2.0
v0.1.1
v0.1.0
Added
- (tui) crimson look, context footer and subtle motion; abandon a starting run without quitting (#8)
- add a terminal UI (#7)
- train on Runpod GPU pods (#6)
- (cli) add train, train attach, train cancel and runs ls
- (runs) start, follow and cancel training runs
- (runs) add run IDs, run.json records and metric summaries
- (events) add training metric and job status events
- (exec) run jobs over SSH with tar transfers
- (exec) build container and native job commands
- (exec) run jobs locally in their own process group
- (exec) add the Executor trait, job scripts and line reading
- (train) render the Axolotl config, commands and artifacts
- (train) add the metrics plugin and parse the lines it writes
- (train) write the Axolotl config as YAML that PyYAML reads back
- (config) add training hyperparameters and target runtime options
- (llm) support budgeted extended thinking for Claude 4.5 models
- (cli) add the pipeline commands with progress and cost summary
- (pipeline) add the answers and split stages
- (pipeline) add the subtopics and questions stages
- (events) add the stage event bus
- (dedup) add lexical and embedding deduplication
- (prompts) add overridable prompt templates and write them on init
- (pricing) read model prices from the provider listing
- (llm) add the anthropic protocol and build clients from config
- (llm) add the openai protocol client with reasoning extraction
- (llm) add client trait, completion types, errors and retries
- (dataset) add records, stable IDs and crash-safe JSONL files
- (config) add per-role request settings and pipeline tunables
- (cli) add init and config check commands
- (secrets) resolve literal and vault KV v2 secret references
- (config) load layered settings and reject env-only keys in file
- (config) add typed settings and semantic validation
Changed
- (exec) stop re-exporting CONTAINER_ROOT
- (pricing) read model listings in one place
Documentation
- say where a container engine keeps the Hugging Face token
- (runs) say how many failures MAX_FAILURES really allows
- (exec) state what environment a job starts with
- correct the cancel and retry wording
- document training, targets and runs
- (llm) reword the thinking configuration comment
- (dataset) state what the ID separator guarantees
- say that split does not accept --force
- document the data pipeline, configuration and releases
- fix precedence, vault example and secret handling notes
- document setup, configuration and secrets
Fixed
- (templates) pin the example docker image by digest
- (config) keep an out-of-range integer a string, not a float
- (exec) signal a job's process group under every sh
- (runs) drain the whole metrics file when the job ends
- (exec) validate a job's secrets on every target
- (exec) keep an interrupted cancel from mislabelling a finished run
- (cli) cancel a run in any state that still has a job
- (cli) load the configuration once and explain what cancel needs
- (runs) report a cancel that succeeded even when its record cannot be updated
- (runs) cancel untracked jobs and keep logs of cancelled runs
- (runs) validate run IDs on every path and skip unreadable records
- (exec) align missing run dir downloads and harden SSH tests
- (exec) explain why a cancel can find the group still present
- (exec) cap the tar error text
- (exec) drain tar stderr, keep secrets out of local jobs and guard nested copies
- (exec) let cancel markers take priority over the exit code in status
- (exec) report cancelling jobs as running and never stall on long lines
- (exec) make cancel and status safe against finished and recycled jobs
- (train) refine the template warning, eval cadence and copy errors
- (train) create the metrics directory before the plugin writes to it
- (train) escape YAML line-break characters in quoted strings
- (config) keep non-canonical integers in axolotl_extra as strings
- (config) reject empty target image and venv values
- (config) give env values of axolotl_extra their types
- (config) restrict target paths and image references to safe characters
- (dedup) return ready futures from the lexical deduplicator
- (dataset) read JSONL files as bytes
- (pipeline) warn when every usable answer is orphaned
- (pipeline) spread the eval set over topics, then subtopics
- (pipeline) log the provider error when saving after a stop fails
- (config) say "at line N, column M" in TOML syntax errors
- (config) reject a temperature together with thinking
- (config) cap pipeline.concurrency at 1024
- (pipeline) leave orphaned answers out of the split
- (pipeline) size the eval set from all usable answers
- (pipeline) never train reasoning from parents that hide it
- (dedup) embed in chunks and skip seeding topics with nothing to fill
- (dataset) keep a valid last line that only lost its newline
- (cli) resolve secrets lazily and connect each role once
- (pipeline) train only raw reasoning and keep finished answers on fatal stop
- (pipeline) resume partial topics and isolate per-item dedup failures
- (pipeline) report an item's final failure exactly once
- (pipeline) match a JSON array past bracketed prose
- (dedup) retry embedding requests under the pipeline's retry policy
- (config) never echo TOML source in syntax errors
- (config) keep offending values out of type error messages
- (config) reject NaN max_hours and split validation into focused checks
- (cli) honor -C in init and merge an existing .gitignore
- (logging) keep dependency error lines out of default output
- (secrets) keep vault error sources and harden token lookup
- (cli) never print .env content on a parse error
- (config) reject log in overbrainer.toml
- (config) accept env overrides of numeric target fields
- (config) replace load's generic hasher with an EnvSource enum