Repository navigation
Integrate practical native MLX SFT and bounded GRPO candidate - #277
Merged
Merged
Conversation
ProfSynapse
marked this pull request as ready for review
October 27, 2026 16:31
This was referenced Oct 27, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds faithful conversation SFT consumption on the Mac, reproducible dataset preparation/token profiling and a separate bounded native MLX GRPO implementation through the existing public TrainingAPI/provider-family/guardian path.
Both practical SFT datasets completed one 100-row training pass (50 updates, batch two), followed by saved-adapter reload evaluation on all 20 held-out rows. Independent review authenticated both signatures, ten delivered artifacts, exact token/mask projections and nonzero changes to all 40,960 adapter parameters. SFT ML evidence remains bound to
d57dc55e.The separately admitted native GRPO attempt at
ad17275anow succeeds: three actual updates, three varied reward groups/48 calls, 102 generated tokens, all 40,960 saved parameters changed, independently recomputed delta L2 0.0407861391343181, and finite saved-adapter reload replay objective. Public verification and all five public role downloads/hash checks pass. The prior failed19768026attempt remains consumed and unclassified. Exact evidence and limits.GRPO uses an explicit named profile, an attributed on-policy objective, bounded declarative rewards and generated-token masks. Pinned Transformers list IDs are requested explicitly. The tokenizer-only example retains native POSIX custody, verifies exact publication commitments and refuses every unexpected direct entry, including empty directories. Both actual Mac positive and empty-directory negative checks pass.
Validation:
ad17275a; 45 focused CPU controls and four independent boundaries passed. Earlier integrated checks include 133 CPU controls and seven CURRENT closure/lock maintainers.c6bc8067preserves all 744 qualified captured source members. Final wheel executable/runtime assets equal the qualified source; only README metadata and RECORD differ. Wheel/sdist source-byte checks and the actual fresh installed base-only consumer pass.c6bc8067and execution evidencead17275a, including signature/selection, all captured source members, five delivered hashes, independent tensor/reward checks and final package equivalence. All 18 current CI checks passed on exactc6bc8067; final independent review and package/installed-consumer gates are clear.This is restricted OSS engine integration with exact-fixture evidence. Ordinary native qualification, public logs/cancel/reconcile, broad lifecycle reliability and full desktop distribution remain gated. GRPO replay checks saved-adapter consistency; it is not held-out quality or practical tool-use evaluation. No broad stable-v1 promotion is claimed.