README.md

metadata

pretty_name: ClimbMix Scored — Concept-Probe Detections (shards 25-49)
tags:
  - interpretability
  - probing
  - gemma-2
size_categories:
  - 1B<n<10B

climbmix-scored-overflow

Concept-probe detection scores for ClimbMix-shuffle shards 25-49 (nanochat training data), gemma-2-2b layers 6/8/14 co-located per token, FULL TOKEN COVERAGE: every parquet row appears in docs_<sid>.jsonl, no min-length filter, no truncation — docs longer than 2048 gemma tokens are scored in consecutive non-overlapping 2048-token windows (window-boundary context truncation is an accepted artifact), so sum(n) equals the shard's total gemma token count. Probes and quantization/standardization are FROZEN from corpus-scores (calibrated once on shard 320 — never refit), so the two datasets are directly comparable (note corpus-scores itself is first-2048-window only). Shard->repo map: assignment.json (25 shards/repo, strictly in order across 8 repos). No DoM scores in this dataset family.

• scores_<sid>.npy int8 [n_tokens, 3, 54] — axis1: 0=L6, 1=L8, 2=L14; axis2 = concept per columns.json. Dequant: score = int8*scale[l][c]+zero[l][c].
• tokens_<sid>.npy int32 — full BOS-free gemma ids, docs in parquet order
• docs_<sid>.jsonl {doc, start, n} — n = full doc token count (0 allowed)
• Scoring: gemma-2-2b bf16 EAGER attention; probes from concept-probes-gemma2-2b gold_probes.