Datasetpublic

abotresol/neutral-vectors-gemma-4-31b-it-postfix

Emotion vectors, google/gemma-4-31b-it (corrected extraction) Residual-stream activations for google/gemma-4-31b-it, pooled per story and averaged per emotion. Each emotion ends up as one direction in the model's activation space. Read LINEAGE.md before using this. This set supersedes abotresol/neutral-vectors-gemma-4-31b-it. The earlier extraction ran while the tokenizer padded on the left, so the step that skips a story's first 50 tokens skipped padding instead. This set… See the full description on the dataset page: https://huggingface.co/datasets/abotresol/neutral-vectors-gemma-4-31b-it-postfix.

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes515downloads

Dataset Card

Emotion vectors, google/gemma-4-31b-it (corrected extraction)

Residual-stream activations for google/gemma-4-31b-it, pooled per story and averaged per emotion. Each emotion ends up as one direction in the model's activation space.

Read `LINEAGE.md` before using this. This set supersedes `abotresol/neutral-vectors-gemma-4-31b-it`. The earlier extraction ran while the tokenizer padded on the left, so the step that skips a story's first 50 tokens skipped padding instead. This set re-extracts the same corpus with padding forced to the right. LINEAGE.md gives the measured before-and-after impact. The predecessor stays up, unmodified, as the "before" side of that comparison.

What is in it

Path	Contents
`<emotion>/layer_<N>_resid.npy`	the mean vector for one emotion at one layer
`emotion_vectors.json`	the same per-emotion means, in the reference's format
`shards/<emotion>__<idx>.npy`	one `[layers, d_model]` array per story, before averaging
`manifest.jsonl`	per-story metadata: emotion, index, text SHA-1, token count
`run_config.json`	the full extraction configuration
`LINEAGE.md`	what this set corrects, and by how much

Shards are published so the per-emotion means can be recomputed, resampled or subsetted without running the model again.

How it was made

—Corpus: `abotresol/neutral-transcripts-gemma-4-31b-it`, transcripts written to carry no emotion
—Layers: every third, 0 to 57
—Pooling: mean over non-padding tokens after position 50, stories truncated at 512 tokens. The first 50 tokens are dropped as narrative framing, which is the convention the source paper used.
—Precision: bf16 weights, fp32 activations. Seed 20260720.
—Across stories: a token-weighted mean, so a long story counts for more than a short one.

Extracted by scripts/extract_emotion_vectors.py in gemma4-emotion-vectors, adapted from sinievanderben/emotion_experiment.

Reproducing

Re-extracting needs the model weights and a GPU with enough memory for a 31B model in bf16. Analysis does not: the per-emotion means in this repository are enough to redo the geometry and detection work on a laptop.

uv run python scripts/extract_emotion_vectors.py

Caveats

A 2-3 day research sprint, not a reviewed publication. The write-up (https://github.com/Antonio-Tresol/gemma4-emotion-vectors) records which findings survived a falsification pass and which did not. These vectors describe a model reading emotions in text; that is a different claim from the model having them.

Licence

MIT, matching the project repository. The model weights and the story corpora carry their own licences.