Repository navigation
Releases
v2.7.0
What's Changed
- Checkpoint eval defaults: default task is now
evals/kmmlu_redux.pyacross SFT, DPO, OPD, SDFT and AsyncGRPO (GRPO usesinspect_evals/math);DEFAULT_EVAL_TASKSand docs updated to match - AsyncGRPO defaults restructured with section headers; eval still runs on
eval_devices: "2"since the rollout server sits on GPU 1 - OPD KMMLU recipe: constant LR with warmup, 5 epochs
- New ZeRO-2 and ZeRO-3 CPU offload accelerate configs
- Removed redundant
qwen3-0.6B-sftconfig andconfigs/DPO/qwen3-0.6B-dpo.yaml
Each release carries a CycloneDX SBOM (sbom-2.7.0.cyclonedx.json + .sha256) generated from uv.lock.
Full Changelog: v2.6.0...v2.7.0
v2.6.0
dna-factory 2.6.0 renames on-policy distillation to OPD (On-Policy Distillation) in the code, configs and docs. Training behavior is unchanged.
What's Changed
distill.pyis nowopd.py, with configs inconfigs/OPD/(_defaults-OPD.yaml,*-opd*.yaml),DnotitiaOPDTrainerindna_factory/dnotitia_opd_trainer.py,docs/opd.mdandutils/measure-opd-kl.py. (#45)output_dir: autonow tags OPD runsOPDinstead ofDISTILL.- The YAML section header and the startup log header are now
DNOTITIA TRAINER ARGUMENTSfor every entry point, and the unusedTrainingSpec.trainer_typefield is removed.
Full Changelog: v2.5.2...v2.6.0
v2.5.2
Docs-only patch release on the 2.5 line. The code is the same as v2.5.1.
Docs
- The README intro now covers all five trainers, including SDFT (Self-Distillation Fine-Tuning).
- On-Policy Distillation (
distill.py) is now called OPD in the README anddocs/. Paper and blog titles keep their original names. - The README news bolds only the acronyms for GRPO and DPO.
- Added NIA to the README acknowledgments.
Supply chain
-
sbom-2.5.2.cyclonedx.json(CycloneDX) and its.sha256checksum are attached. Verify the download, then scan it for known vulnerabilities:sha256sum -c sbom-2.5.2.cyclonedx.json.sha256 grype sbom:sbom-2.5.2.cyclonedx.json
Full Changelog: v2.5.1...v2.5.2
v2.5.1
Patch release on the 2.5 line. It is the first release that ships a CycloneDX SBOM. The code is the same as v2.5.0; see that release for the new features (SDFT, the unsloth backend, Qwen3.8-Flash-Next LoRA SFT, KMMLU recipes, and the TRL 1.17.1 / transformers 5.22.1 bump).
Supply chain
-
SBOM on every release:
sbom-2.5.1.cyclonedx.json(CycloneDX 1.7) lists all 300 packages pinned inuv.lock, including theunslothextra. Its.sha256checksum is attached alongside it. Verify the download, then scan it for known vulnerabilities:sha256sum -c sbom-2.5.1.cyclonedx.json.sha256 grype sbom:sbom-2.5.1.cyclonedx.json
-
Releases in this repo are immutable, so the SBOM is now attached while the release is still a draft. Pushing a version tag creates the draft. The steps are in
docs/release.md. v2.5.0 was published before this change and has no SBOM.
Docs
- Shortened the SDFT comment in the README quick start.
Full Changelog: v2.5.0...v2.5.1
v2.5.0
dna-factory 2.5.0 adds two new training methods (SDFT and an opt-in unsloth backend), Qwen3.8-Flash-Next LoRA SFT, and KMMLU recipes for GRPO and on-policy distillation. It also bumps TRL to 1.17.1 and transformers to 5.22.1.
New Features
- SDFT (Self-Distillation Fine-Tuning): new
sdft.pyentry point. The teacher is the model itself, shown a per-rowprivileged_contextthat the student does not see. It trains on GSM8K demonstrations by default. Seedocs/sdft.md. (#43, @cm8908) - unsloth backend: opt-in with
use_unsloth: truefor SFT, DPO, GRPO and distillation (uv sync --extra unsloth). Single GPU only. Seedocs/unsloth.md. (#34, @cm8908) - Qwen3.8-Flash-Next LoRA SFT, plus hardening of its training path. See
docs/qwen38-flash-next.md. (#40, @cm8908) - Faster qsa indexer: vectorized across the batch and its padding modes. (#42, @cm8908)
- Thinking-model GRPO rewards that close accuracy leaks (
dna_factory/rewards/reasoning.py). - KMMLU recipes: GRPO with multiple-choice rewards, and on-policy distillation of Qwen3-4B from Qwen3-32B.
Changes
- GRPO defaults: halved default completion and eval lengths,
vllm_max_model_lengthcapped at 12480, andvllm_importance_sampling_correctionset explicitly. - Distill defaults:
vllm_max_model_lengthcapped. - GRPO config variants (no vLLM, server mode, DAPO, Dr. GRPO, weighted rewards, judges, unsloth) are now documented as snippets in
docs/grpo.mdand no longer shipped as separate files. (#39, @cm8908) - Dependencies: TRL 1.17.1 and transformers 5.22.1 (obsolete workarounds removed), plus bumps to deepspeed, wandb, kernels, ruff and others. (#44)
Tests
- Unit tests for
resample_by_weight. (#37, @chloeseo0312)
Supply chain
- This release has no SBOM attached: releases here are immutable, and the upload ran after publishing. The same code ships with an SBOM as v2.5.1.
Full Changelog: v2.4.0...v2.5.0
v2.4.0
What's Changed
- lint: fix remaining ruff findings and pin the rule set by @likejazz in #24
- refactor: streamline file handling and improve code clarity by @likejazz in #25
- fix: sync periodic checkpoint saves across ranks by @likejazz in #26
- feat: add AsyncGRPO execution mode by @cm8908 in #28
- feat: evaluate every checkpoint with Inspect and log the scores to W&B by @likejazz in #30
- fix: parse repeated and valueless --config arguments 🤖🤖🤖 by @Momoyeyu in #32
- feat: honor dataset weights in DPO and GRPO 🤖🤖🤖 by @Momoyeyu in #33
Full Changelog: v2.3.0...v2.4.0
v2.3.0
What's Changed
- Fix "ZoRO-3" typo in README Multi-GPUs section by @likejazz in #13
- docs: add a Development section (ruff + pytest) to README by @likejazz in #15
- fix: fail when --config points to a missing YAML by @HarshRajSinghania in #14
- docs: consolidate README intro and describe the four trainers by @likejazz in #17
- ci: add GitHub Actions workflow by @Yyash64 in #16
- ci: fix format check failure and polish CI workflow by @likejazz in #18
- utils: add dataset validator for prompt-only (GRPO/Distill) datasets by @likejazz in #20
- ci: enable the lint gate and tidy the workflow by @likejazz in #21
- ci: track master after default branch rename by @likejazz in #22
New Contributors
- @likejazz made their first contribution in #13
- @HarshRajSinghania made their first contribution in #14
- @Yyash64 made their first contribution in #16
Full Changelog: v2.2.0...v2.3.0
v2.2.0
What's Changed
- feat: add On-Policy Distillation support (distill.py) by @Matlab7 in #1
- Add dynamic sampling for zero-advantage GRPO groups by @cm8908 in #2
- fix: stale logging and output-directory tests by @cm8908 in #3
- fix(grpo): make dynamic sampling fire at pd > 1 and remove the wrong fallback by @cm8908 in #4
New Contributors
Full Changelog: https://github.com/dnotitia/dna-factory/commits/v2.2.0