Manual full pipeline

Research-grade control — no automated search.

Phase A — Data#

1. Generate 128–512 harmful/harmless pairs (or use public refusal datasets)
2. Filter — run base model; drop pairs where both labels agree (no signal)
3. Store data/paired_v1.jsonl + data/eval_v1.txt

Phase B — Directions#

1. Implement activation harvester (see ../methods/manual-transformers-pipeline.md)
2. Hook: post-MLP residual per layer
3. Token index: last prompt token (document choice)
4. Save runs/<run_id>/directions.pt + metadata.json

Phase C — Edit#

1. Start with layers 0.45n … 0.65n of depth
2. Apply MLP down_proj ablation α=0.5
3. Save candidate model_a/

Phase D — Evaluate#

Phase E — Iterate#

• If under-ablated: ↑ α or add layers
• If capability hit: ↓ α or shrink layer band
• Log every candidate in runs/RESULTS.md

Phase F — Publish artifact#

abliterated-<base>-<date>/
  README.md          # method, eval, ethics notice
  config.json
  model.safetensors