Repository navigation
Prepare representative native MLX SFT conversation datasets - #275
Closed
ProfSynapse wants to merge 2 commits into
Closed
ProfSynapse wants to merge 2 commits into
ProfSynapse wants to merge 2 commits into
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The native SFT fixture used only three train rows and two validation rows. This layer adds reproducible config-first preparation for 100 train/20 held-out public HF conversations and 20/4 explicitly positive historical Nexus tool examples per manager, preserving source commitments, original splits, complete context and final-target supervision through the existing immutable conversation normalizer.
The host example performs bounded public parquet acquisition, audited eligibility and grouped selection, and a tokenizer-only profile that retains every row ID, token, label, generation scaffold and actual native padded batch membership under a 16 MiB private output bound. Two canonical training inputs and review-only proposals select sequence lengths 512/3072, batch 2, 50 updates for one complete 100-row pass, all 20 validation rows, and a 300-second worker limit. They grant no execution authority or throughput/model-quality claim. The Nexus projection preserves historical serialized
useToolstext; current Nexus API behavior is not qualified.Stacked on PR274's independently accepted native host adapter. No runtime, worker, model, dependency lock or admission changes in this layer.
Validation: 90 focused CPU tests pass (new selection/profile tests, conversation preparation and both host suites); Ruff and diff checks pass; canonical skills synced and checked. Both genuine Linux publications independently verify and selection reproduces byte-for-byte. Mac tokenizer-only diagnostics pass HF at 2048 (max 305) and Nexus at 8192 (max 2796); the earlier 2048 Nexus rejection and incomplete-config reports remain retained. Final selected 512/3072 native token profiles and independent source review are pending root execution from this pushed source, before any capture/admission/training. No MLX/model-weight loads or training effects performed by this task.
Integrated into merged PR #277, merge
5c9a5cf9a4b19e5ce193c3c09a46df22c86296a7. Its reviewed/tested tree includes this work and the final Mac qualification fixes. This draft is superseded; branch and retained evidence are preserved.