Skip to content

Latest commit

 

History

History
39 lines (26 loc) · 2.1 KB

File metadata and controls

39 lines (26 loc) · 2.1 KB

Phase 0/1 targeted data recovery

This document applies the pending audit/update notes that were not previously landed cleanly.

Source of update

The pending note identified four concrete problems:

  1. The README undersold the real incumbent by still referencing older v6-era status.
  2. The repo had packaging drift around .env.direct-runtime.example.
  3. Stub-like distillation records were present and should not be treated as useful training signal.
  4. Phase 1 should build targeted examples for weak capabilities instead of adding generic data.

Phase 0 cleanup applied

  • Added .env.direct-runtime.example from env.direct-runtime.example so validators that expect the dotted file can pass.
  • Scanned dataset JSONL files for the known stub signature:
    • Seed distillation task ...
    • Bounded structured response ...
  • Removed the matching stub rows from datasets/distillation_sft/seed_records.jsonl.
  • Archived removed rows in reports/production_audit/phase0_removed_stub_records.jsonl.

Note: the pending note referenced 385 stubs from an earlier analysis. In this uploaded package, the directly detectable stub signature appears in fewer rows. This patch removes the stubs actually present in the uploaded package rather than inventing a larger deletion count.

Phase 1 targeted data added

Added 150 v2-candidate-only examples:

File Capability Records Purpose
datasets/distillation_sft/phase1_instruction_following.jsonl instruction_following 50 exact formatting, constraints, patch discipline
datasets/distillation_sft/phase1_continuity.jsonl continuity 50 preserve incumbent/candidate/dataset doctrine across turns
datasets/distillation_sft/phase1_reflection.jsonl reflection 50 critique overclaims, correct attribution, revise precisely

These files are not incumbent data. They are staged for a future v2 candidate training run.

Required next step

A future candidate should be trained with these files, benchmarked, scored, and passed through Gate v2. Until then, these records are corpus improvements only; they are not model improvements.