Skip to content

Releases

relm 0.6.1

Choose a tag to compare

@Vadale Vadale released this 23 Oct 19:53
cbecff9

relm 0.6.1 is a maintenance release for Python reference tooling and shared desktop builds.

  • Update maintainer-tooling urllib3 to 2.8.0 and the optional CPU Hugging Face environment to PyTorch 2.13.0; remove unused Accelerate. Python is not required to use relm from R. Frozen numerical references and their producing manifests remain unchanged.
  • Correct extendr 0.9.0 R variable lookups, including inherited bindings, missing versus NULL values, and safe containment of promise and active-binding errors. Developer and offline package sources use the same verified patch.
  • Fix strict source verification under Windows line-ending settings and retain the supported Linux, macOS and Windows x64 checks.
  • Document deferred Windows ARM64 and WebAssembly work, and remove two demonstrated duplicate CI invocations while preserving numerical, security and process boundaries.
  • Admit 0.6.1 explicitly in the batch/service recipes without admitting unknown future versions.

Windows ARM64, WebAssembly and CUDA investigation remain deferred. This release does not claim browser inference or a CUDA repair. The published 0.6.0 tag/archive and the pending 0.5.0 CRAN submission are unchanged; no duplicate CRAN submission is made.

Validation and distribution

The twelve required main checks and the documentation deployment passed for commit cbecff997eb384803894aa8e47ff107b8e58fe3a. The published source archive and eleven supported desktop binaries were downloaded and checked against their published SHA-256 identities. Supported r-universe desktop package checks finished with Status: OK; opt-in tests and binary vignette re-execution remain explicitly skipped where configured.

A fresh R 4.6.1 session on Linux x86_64 installed the published binary into an isolated library with the existing nanoarrow 0.9.0 dependency. Package identity, 16 public exports, 33 registered calls, all 250 R functions against the published source, and model-free loading/argument boundaries passed. No model download, inference or native rebuild was performed. This is not a clean-VM, RStudio, Mac mini or browser-inference claim.

All five Python-tooling Dependabot alerts addressed by this maintenance release are now marked fixed by GitHub. The external r-universe Wasm job still fails because its required target runtime path is not configured; Windows ARM64 and WebAssembly remain deferred. The overall external build is therefore not all-green, despite successful supported desktop publication.

Published source: relm_0.6.1.tar.gz, also attached unchanged to this release.

  • Source commit: cbecff997eb384803894aa8e47ff107b8e58fe3a
  • Source archive size: 8,628,327 bytes
  • Source archive SHA-256: ef2a67dc1d1d7046ee8e975999c150e638e2a2d5dca64969f3eb844484bf6b17
  • Linux x86_64 / R 4.6.1 binary SHA-256: 5e72e437d40cee16d20338d30f344e1c2f14fd50a61647081f6abfc7e0d5f9c5

The pending CRAN submission is unchanged and remains awaiting feedback. This release does not replace that submitted archive.

relm 0.6.0

Choose a tag to compare

@Vadale Vadale released this 23 Oct 02:45
8f6cb33

relm 0.6.0 is published on r-universe.

llm(..., threads = NULL) adds an optional CPU thread count while preserving the existing default and R check-time limits. This release also recognizes the exact supported modern Qwen/Gemma chat templates, including the pinned Qwen3.5-35B-A3B MoE template. Unknown modern templates remain explicit refusals. Recognition does not guarantee model accuracy or speed; see the model guide.

The maintainer reference-tooling pin is updated to certifi 2026.10.19. Existing golden outputs and numerical bounds are unchanged. GitHub now reports the associated critical certifi advisory as fixed.

Release source: 8f6cb335869323058d54207c45bebf79630e0015.
The published source archive is 8,621,377 bytes, SHA-256 4f20bce48ccd0603fa41fe021123fcf68ec17c3080f268dca1e7aa5a784ba2a3.

All twelve required ordinary checks and the corrected scheduled gates were accepted in their recorded scopes. The r-universe build published eleven verified binaries: Linux x86_64/ARM64, macOS x86_64/ARM64, and Windows x86_64 across the listed R versions. Their actual package checks passed. The public Linux R 4.6.1 binary was installed and its public boundaries checked in a fresh R process and isolated library, without rebuilding native code or replaying model studies. This is not a clean-VM or IDE-specific acceptance claim.

Windows ARM64 and webR/WebAssembly builds failed and are not supported by this release's distribution acceptance. The overall r-universe workflow therefore remains marked failed despite successful publication for the platforms above. No failing check was disabled or reclassified as passing. CUDA remains a separately validated opt-in source-build scope; the previously unreproduced CUDA issue is deferred, not claimed fixed.

The tested source archive retains its prepublication candidate wording in README/NEWS. This dated release announcement records publication of those exact tested bytes and supersedes that earlier distribution status.

The existing 0.5.0 CRAN submission remains unchanged and pending. This release does not submit a replacement or announce 1.0.

relm 0.4.0

Choose a tag to compare

@Vadale Vadale released this 12 Oct 10:38
3bd5da7

relm 0.4.0

Released from commit 3bd5da7399dc163d8c9e4a4e1865a875e7186e71, whose tree exactly matches the verified packaging candidate. Maintenance and final packaging are integrated through PR65 and PR66.

This release adds component projection steering, recorded contrast directions, live state observation and intervention graphics, token streaming and the complete research workflow. Maintenance simplifies shared validation and condition handling, strengthens handle provenance and excludes native fixture hooks from production builds. The README and simplifier compendium describe the supported path and code responsibilities.

Production packaging now uses the correct host R executable and explicit C++ linkage. A reversible diagnostic-only vendor patch removes compiler-warning suppression; it does not change the pinned upstream version or numerical calculations. Demo A expresses its existing small-fold metric policy and collapsed intervals without warnings.

The actual Mac production source check reports zero errors, zero warnings and one NOTE, identifying a new CRAN submission and the 5.15 MB archive. It passed 5,010 expectations with 130 explicit optional/fixture/network skips; examples, six vignettes and PDF/HTML manuals passed. Fresh production installation has no native fixture hooks. Prior actual RStudio README and demo acceptance retains its parent source/DLL provenance; warning corrections were checked without replaying model inference.

The projection efficacy experiment remains a modest, truncated fixed-corpus result, not a generalized quality claim. Default numerical behavior is not asserted equivalent to the F64 oracle; earlier cross-schedule outliers remain disclosed. Windows/CUDA and CRAN submission are outside this release. See the release report for exact source scopes, limits and artifact hashes.

All nine final CI jobs passed: four R checks and five Rust/integrity checks, including full default, no-spill and release-mode execution. Ancillary statistical-fixture warnings and nonblocking artifact ZIP diagnostics are retained in the PR; unsupported ZIP entries are not claimed archived. Package checks themselves passed.

Install

install.packages("relm", repos = c(
  "https://vadale.r-universe.dev",
  "https://cloud.r-project.org"
))

Fresh r-universe binary installation is verified on macOS Apple Silicon with R 4.5.1. The downloaded binary and installed RemoteSha match release commit 3bd5da7399dc163d8c9e4a4e1865a875e7186e71; all 49 installed files match the binary payload. One cached CPU model load and one short generation returned Ready, with the model closed and no production fixture hooks. This was a new library and R process on the existing Mac, not a clean machine/VM or a model-quality evaluation.

The Mac R-oldrel build, check and binary storage job succeeded with Status: OK. Its binary SHA256 is ee716dfb0d403e3136d5fa84edf406d0e128d5d57e2698f83580cb01168eac1d. It was built with R 4.5.3, so loading it under host R 4.5.1 emitted the retained patch-version warning; the installation and generation passed. The automatic Windows/Wasm jobs failed outside this release's supported scope; the entire r-universe matrix is not presented as passing.

The attached source archive is the checked production package. The attached macOS binary targets Apple Silicon and was built with R 4.5.1; it is not a Windows, CUDA or Intel Mac binary. SHA256SUMS identifies both assets:

  • relm_0.4.0.tar.gz: 48978da212cad853ce56055a8d8733328ce4835053b3d9c12e8e8a9544339048
  • relm_0.4.0.tgz: 971b72086c4bce52170d191e1bcc24af6a7b373c98a9b6d76e5a5855f16c04d2

relm 0.3.0 — structured output, probes and repeatable applications

Choose a tag to compare

@Vadale Vadale released this 30 Sep 08:22
82fcf13

relm 0.3.0 brings structured output, statistical probes and repeatable local applications to the existing base-R model research workflow.

New in 0.3.0

  • Constrained JSON: llm_generate(schema = ...) accepts the supported bounded schema profile and returns validated output or a classed error.
  • Statistical probes: llm_probe() and activations() provide binary ridge probes, explicit source groups, development-only model selection, held-out evaluation and conditional intervals, with S3 summaries, plots and predictions.
  • Native Spark: Spark-X2.5-4B generation through embedded llama.cpp b10828, including its official non-thinking opener for constrained chat. Spark activation tracing remains unsupported.
  • Application templates: offline batch setup/run/resume and a loopback HTTP service with a persistent worker, durable tickets and bounded resources. Prepare fresh 0.3.0 environments; historical development snapshots are not reused.
  • Fixes and documentation: spill-file collision protection, strict trace indexing, explicit CPU device selection, classed argument errors, updated READMEs/install guide and probe reference pages.

See NEWS, quickstart and application examples.

Install

install.packages("relm", repos = c(
  "https://vadale.r-universe.dev",
  CRAN = "https://cloud.r-project.org"
))
packageVersion("relm")

r-universe now serves verified 0.3.0 Mac and Linux binaries from the release commit. It follows main, so check the installed version. The source archive and verified Mac arm64/R4.5 binary are attached with SHA256 checksums. Source builds require R >= 4.5, Rust >= 1.85.0, CMake >= 3.28 and a C/C++ toolchain. Windows/CUDA acceptance remains deferred.

Verification and scope

Distribution verified on 2026-09-30: r-universe build serves 0.3.0 from 82fcf1346f35a3bc866f4c4d8da58fd7d4cd7aa9, with all eight listed Mac/Linux arm64/x86_64 binaries successful at that same source. Mac binaries cover R 4.5.3/4.6.1; Linux binaries cover R 4.6.1/4.7.0.

A binary-only installation into a new library and fresh R process on the founder's Mac passed, including exact package version, RemoteSha, dependency paths and one seeded generation with the cached SHA256-pinned Qwen model. The downloaded Mac arm64/R4.5 archive matches SHA256 f206667f8407f2d0a2537094d4e2d486a13582fab4d857e5159f334a9aeb3154. Local R 4.5.1 reported that relm was built under R 4.5.3; this warning is retained. No source compilation was needed. This was a fresh library on the existing Mac, not a clean-machine/VM test or a new model-quality evaluation.

All nine final PR checks passed at 392ca55b5f24228c6eacf1da4f028aec7f621d63: R CI and Rust CI.

The release report retains full provenance: built-package tests/examples/vignettes/manuals, actual RStudio README and pinned-model demos, 466 batch assertions and 11 service environment cases. Earlier service acceptance includes 1,000 same-worker requests on Mac and Linux; parent-source stress and later lifecycle measurements keep their exact identities and prior failures.

Schema validity and operational reliability do not establish extraction accuracy. The frozen small-model D1 pilot remains 10/10 schema-valid, 2/10 task-valid and 0/10 fully grounded; all four quality promotion gates failed. Predictive decodability does not establish causal use, and interventions are research mechanisms rather than safety guarantees.

CRAN preparation remains Phase 9

Under founder-approved D-035, this GitHub/r-universe release retains two documented CRAN-readiness warnings: vendored C++ pragmas and source builds that obtain locked Rust crates online. No failed test or numerical/resource threshold is waived. No clean CRAN check or CRAN acceptance is claimed. Retained notes concern a new CRAN submission, unavailable time verification and old local HTML Tidy. Model-gated pkgdown examples and historical Demo A warnings are itemized in the release report.

The next default roadmap scope is WP9 asynchronous generation followed by WP10 token streaming. No new work package starts as part of this release.

relm 0.2.0 — vision

Choose a tag to compare

@Vadale Vadale released this 15 Jul 21:32
b8bd7de

relm 0.2.0 adds vision. A handle loaded with a model's mmproj projector takes image input across generation and embeddings — on the same vendored engine as 0.1.0, with the text-only paths byte-identical.

Install (macOS Apple Silicon + Intel + Linux binaries, no toolchain)

install.packages("relm", repos = c("https://vadale.r-universe.dev", getOption("repos")))

Seeing machines

m <- llm(llm_download("qwen2-vl-2b-instruct-q4_k_m"),
         projector = llm_download("qwen2-vl-2b-instruct-mmproj-f16"))

llm_generate(m, "What color is the shape in this image?", images = "circle.png")
#> "Blue"

llm_embed(m, "", images = "circle.png")   # image and text in one vector space
  • llm(projector=) loads the vision tower alongside the model.
  • llm_generate(images=) asks about an image.
  • llm_embed(images=) puts images and text in one space.
  • Registry aliases pin an Apache-2.0 default pair (qwen2-vl-2b-instruct-q4_k_m + qwen2-vl-2b-instruct-mmproj-f16), SHA256-verified against an immutable revision, fail-closed.
  • vignette("vision", "relm") walks the whole path and is self-contained — it draws its own test images with base graphics, so it renders with or without a local model.

How it is tested

The image encoder's raw output is checked value-for-value against an unpatched upstream llama.cpp reference — bit-exact (max |Δ| = 0.0 over 98,304 values) on both macOS arm64 and Linux x86_64. Each runner builds that reference itself, because a float reference belongs to the machine that produced it: the same pristine upstream build disagrees with its own arm64 self across ISAs by far more than any tolerance worth setting.

What that gate does not cover is stated plainly in DECISIONS.md (D-026 fourth addendum): the encoder path carries none of relm's own patch, so bit-exactness there is the expected result. What supports vision across architectures is the byte-exact generation golden and the token-ids pin, both green on Linux against a reference recorded on arm.

Notes

  • Images: JPEG, PNG, BMP. The projector file is engine-trusted input, exactly like the model GGUF — prefer the SHA256-verified registry aliases.
  • Audio is not supported in this release, though the engine's multimodal library carries it.
  • Binaries: macOS arm64 + macOS Intel + Linux. Windows is a later phase — build from source meanwhile (needs Rust + CMake).

Full changes: NEWS.md

relm 0.1.0 — first public release

Choose a tag to compare

@Vadale Vadale released this 08 Jul 23:09
215e533

relm runs open-weight LLMs locally from R and hands you the results as base-R data.frames and matrixes — no Python, no API keys. It embeds a vendored, patched llama.cpp in a Rust core. The differentiator over a plain binding: it opens the model up — trace activations, steer the residual stream, ablate units — mechanistic interpretability from R.

Install (macOS Apple Silicon + Linux binaries, no toolchain)

install.packages("relm", repos = c("https://vadale.r-universe.dev", getOption("repos")))

What's in it

  • Model loading (llm) and tokenization (llm_tokens)
  • Text generation (llm_generate) and next-token distributions (llm_logits)
  • Embeddings (llm_embed)
  • The interpretability core: activation tracing (llm_trace), steering (llm_steer), ablation (llm_ablate)
  • Checksum-verified model download (llm_download)
  • Two runnable demos: topics without Python and the anatomy lab

Notes

  • Every numerical feature is validated value-for-value against an independent reference.
  • Text-only; vision is planned for v0.2.0.
  • Binaries: macOS arm64 (Apple Silicon) + Linux. Windows and Intel-Mac binaries are a later phase — build from source meanwhile (needs Rust + CMake).

See the README and docs/getting-started.md.