Find changed features¶

Locate the layers a derivative changed, check a concept direction, and train a crosscoder to list new, lost and shifted features.

pip install "modeldiffr[hf,tasks]"

1. Find the layers that moved¶

modeldiffr layers Qwen/Qwen2.5-0.5B-Instruct huihui-ai/Qwen2.5-0.5B-Instruct-abliterated-v3 \
  --corpus builtin --dtype float32

1850 tokens from 76 texts (0 skipped)
layer      CKA   cosine  rel change  mean shift  norm ratio
    0   1.0000   1.0000      0.0000      0.0000      1.0000
  ...
   14   0.9871   0.9802      0.1543      0.0611      0.9930
  ...
largest change at layer 14

The numbers above show the format; yours depend on the pair. --corpus takes builtin (the quick suite's texts), a .jsonl file with a text field (--field to change it), a .txt file with texts separated by blank lines, or hf:DATASET[:CONFIG]:SPLIT to stream from the Hub.

2. Check a direction you care about¶

modeldiffr directions BASE DERIVATIVE --positive harmful.txt --negative harmless.txt --layers 8-20

Column	Meaning
`cosine`	Agreement between the two models' directions
`base sep`	Separation of the two prompt sets along the base direction, in the base
`deriv sep`	The same separation, in the derivative
`ratio`	`deriv sep / base sep`; well below 1 means the concept was weakened

Use your own prompt sets. Modeldiffr does not ship harmful prompts.

3. Train a crosscoder¶

modeldiffr features BASE DERIVATIVE --layer 14 \
  --corpus hf:HuggingFaceFW/fineweb-edu:sample-10BT:train --texts 20000 --max-tokens 200000 \
  --dict-size 16384 --k 64 --steps 20000 --dtype bfloat16 --device cuda \
  --out audit --report audit/report.json

features id mdrf1:a77a8e2a...
layer 14: 41 new, 12 lost, 230 shifted, 15233 shared, 868 never active
variance explained base 0.871, derivative 0.866
report: audit/features.html

Writing into the audit folder with --report links the two, and the registry then shows the feature counts on the audit's page.

Setting	Small model on a CPU	7B on one GPU
`--max-tokens`	20,000	200,000 to 1,000,000
`--dict-size`	2,048	16,384 to 65,536
`--k`	32	64 to 128
`--steps`	2,000	20,000 or more
`--batch-size`	1,024	4,096

Activations need tokens × hidden size × 8 bytes of memory: 200,000 tokens of a 3,584 wide model is 5.7 GB.

Add --variant delta for a narrow fine tune, and --label-api URL --label-model NAME for automatic labels from any OpenAI compatible server (the key comes from MODELDIFFR_API_KEY or OPENAI_API_KEY).

4. Read the page¶

features.html shows the class counts, a histogram of where each feature's decoder norm lives, fit statistics, and for each class a list of features with the contexts where they fire most. Treat a derivative_only feature with a clear pattern of examples as a lead to test with causal checks, not as a conclusion.

5. Verify¶

modeldiffr verify audit/features.json

PASS  id                         mdrf1:a77a8e2a...
PASS  result_hash                results match their hash
PASS  crosscoder.safetensors     matches its hash
PASS  classes                    16384 features classified

See Representational layer for the method and its limits.