llm-abliteration workflow

Manual measure → analyze → ablate pipeline from jim-plus/llm-abliteration v1.2. NousResearch/llm-abliteration last push 2025-11-27 — treat as frozen; do not use it as the handbook path.

Clone#

bash

git clone https://github.com/jim-plus/llm-abliteration.git tools/llm-abliteration
cd tools/llm-abliteration
pip install -r requirements.txt

bash

python measure.py -m <path_to_model> -o directions.pt

Custom datasets (local .txt, .jsonl, .parquet):

bash

python measure.py -m <model> -o directions.pt \
  --data-harmful ./data/harmful.txt \
  --data-harmless ./data/harmless.txt

Flag	Effect
`--quant 4bit`	On-the-fly bitsandbytes quant for measurement (low VRAM)
`--projected`	Projected abliteration (orthogonalize vs harmless dir)
`--deccp`	Add AUGMXNT/deccp topics to the harmful set. That repo is a Qwen2 PoC + dataset, not a peer CLI.

bash

python analyze.py directions.pt -c

Use charts to pick middle-to-late layers as refusal direction sources.

Edit example YAML (see repo), then:

bash

python sharded_ablate.py abliteration_config.yaml

Flag	Effect
`--projected`	Biprojected measurement
`--normpreserve`	Norm-preserving ablation

bash

python chat.py -m <abliterated_model_path>
python compare.py -a <base> -b <abliterated>

Pair --quant 4bit measure with sharded_ablate.py when Heretic OOMs:

bash

node scripts/fetch-docs.mjs
# reads sources/fetched/llm-abliteration-readme.txt