Repository navigation
feat(assemble): replay sources survive the verdict gate - dataset v2 - #142
Merged
Merged
Conversation
The v1 MR6 dataset shipped with ZERO replay records: the failed- trajectory gate (MR5 lesson, aimed at rollout records) silently ejected the whole teacher corpus (ctf-big, ctf-distill, ctf-first*), and the anti-forgetting mix the consolidation cycle needs never made it into the training file. The retrain then re-created exactly the forgetting failure the replay exists to prevent. - assemble_sft_mr6.py: --replay-source (repeatable). Replay records skip the failed-trajectory gate AND the per-id cap (teacher- namespace IDs), every other gate still applies: contamination (2 eval-leaked ctf-lictor-uaf_heap-001 teacher records dropped, exit nonzero), class-fact, excluded targets, empty share. - run_final_bench.sh: the v2 file joins TRAIN_IDS - the derivation drops uaf_heap-002 from the eval automatically (the sacrifice ledger rides the dataset), eval set after v2 = 10 challenges. - datasets/eval/README.md: the v2 section documents the contract. TDD: test_replay_source_records_kept_without_verdicts (red: the gate dropped all 5), test_replay_sources_skip_the_per_target_cap, test_replay_source_eval_contamination_still_refused (red: the gate let the eval leak through). All green after the fix.
7 tasks done
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Dataset v1 shipped with ZERO replay records: the failed-trajectory gate silently ejected the whole teacher corpus (ctf-big, ctf-distill, ctf-first*) and the anti-forgetting mix never reached the training file. This adds
--replay-sourcetoscripts/assemble_sft_mr6.py(skips the verdict gate + per-id cap; every other gate still applies), and wiresdatasets/sft-mr6-v2/train.jsonlintoscripts/run_final_bench.shTRAIN_IDS so the clean derivation drops uaf_heap-002 from eval automatically (sacrifice ledger rides the dataset).Also: 2 eval-leaked teacher records (
ctf-lictor-uaf_heap-001in ctf-first.jsonl/batch1) are dropped by the contamination gate - replay is not exempt.Tests (mandatory)
TDD red-green:
test_replay_source_records_kept_without_verdicts(red first: the gate dropped all 5 replay records),test_replay_sources_skip_the_per_target_cap,test_replay_source_eval_contamination_still_refused(red first: the gate let the eval leak through). All green after the fix:pytest tests/unit/test_assemble_sft_mr6.py11 passed. Files:tests/unit/test_assemble_sft_mr6.py.Linked Issue
No issue. Scope: MR6 dataset v2 (levier 1 replay) - the consolidation retrain needs the replay mix; without it the retrain re-creates catastrophic forgetting.
Quality Checklist
CHANGELOG.mdunder[Unreleased]