Alexander Whitestone
|
83457cc9a9
|
feat(#691): training pair provenance tracking — source session + model
ProvenanceTracker: added add_provenance(), extract_provenance_from_existing(),
filter_by_provenance(), generate_report() methods.
Fixed save_jsonl() to accept both (path, entries) and (entries, path)
argument orders for backward compatibility.
build_curated.py: every exemplar now gets provenance metadata
(source=curated, source_session_id, model=timmy-curated, timestamp).
Provenance coverage reported in build output.
Acceptance criteria:
- [x] Add metadata to each pair: source_session_id, model, timestamp
- [x] Filter pairs by provenance (exclude_models, exclude_sources)
- [x] Report: pair count by source model
Closes #691
|
2026-04-21 07:14:27 -04:00 |
|
perplexity
|
6507cffc15
|
feat: migrate autolora pipeline into training/
Per direction shift (the-nexus#542).
Replaces the autolora repo (1,500 lines of custom pipeline code)
with config files for existing tools:
- axolotl.yaml: replaces train_modal.py (239 lines)
- mlx-lora.yaml: replaces MLX training scripts
- eval-tasks.yaml: replaces run_eval.py (300 lines)
- Makefile: replaces run_vibes.py, compare.py, convert_to_mlx.py
Data migrated as-is:
- curated_dataset.jsonl (26 gold-standard conversations)
- preference_pairs.jsonl (DPO pairs)
- prompts_vibes.yaml, prompts_nexus_vibes.yaml
- v0-baseline eval results (historical record)
Thin glue kept:
- build_curated.py (data authoring, not infrastructure)
- ingest_trajectories.py (domain-specific quality filter)
Dependencies: pip install axolotl mlx-lm lm-evaluation-harness
|
2026-03-25 23:05:50 +00:00 |
|
