Datasetpublic

kaushikreddyxyz/corpus-scores-dom-layer8

corpus-scores-dom-layer8 Difference-of-means (DoM) STEERING scores for ClimbMix, gemma-2-2b layer 8. Kept separate from the detection score store corpus-scores (overflow). Files (per shard <sid>, ClimbMix shards 320-362, 43 total) scores_<sid>.npy — int8 [n_tokens, 54]; axis 0 token, axis 1 concept (index into columns.json concepts[]) tokens_<sid>.npy — int32 [n_tokens] gemma tokenizer ids docs_<sid>.jsonl — per-document metadata (token spans)… See the full description on the dataset page: https://huggingface.co/datasets/kaushikreddyxyz/corpus-scores-dom-layer8.

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes1.1kdownloads

Dataset Card

corpus-scores-dom-layer8

Difference-of-means (DoM) STEERING scores for ClimbMix, gemma-2-2b layer 8. Kept separate from the detection score store corpus-scores (overflow).

Files (per shard `<sid>`, ClimbMix shards 320-362, 43 total)

—scores_<sid>.npy — int8 [n_tokens, 54]; axis 0 token, axis 1 concept (index into columns.json concepts[])
—tokens_<sid>.npy — int32 [n_tokens] gemma tokenizer ids
—docs_<sid>.jsonl — per-document metadata (token spans)

Provenance

All-DoM steering probes from concept-probes-gemma2-2b gold_probes/layer08 applied to gemma-2-2b residual activations (~46.6M tokens/shard, ~2.0B total). Scores standardized per probe over ClimbMix, then int8-quantized (clipped at ±4σ). Dequantize with dom_quant.json: score = int8 * scale[c] + zero[c].