Solshine/gemma-4-e2b-nla-av-sft-short-labels

av_sft — Hybrid Short-Label Dataset (Opus 4.6 / Sonnet 4.6 / Gemini 2.5 Pro)

Location: data/stage3_v0_1_full_opus_short/av_sft.parquet
Generator: experiments/v8_nla_local/stage2_opus_short_labels.py
Label date: 2026-05-15
Rows: 4,734 (same schema as stage3_v0_1_full/av_sft.parquet + labeler_model column)

Purpose

Tests the label-format-as-lever hypothesis: the original Gemini labels are 200+ char paragraph-length elaborations (e.g., "The model tracks the 'unresolved' status of the 'unresolved' attribute, which..."). Short labels (≤5 words) may force the AV to use the activation signal rather than learn generic template surface area from the long-form labels.

Labeler model breakdown

Labeler	Rows	% of corpus	Notes
opus-4-6	1,361	28.7%	Rows 0–475 + additional Hermes Agent batch (first pass), highest quality per sampling
sonnet-4-6	1,125	23.8%	Rows 862–1,985 (second pass after Opus credits exhausted)
gemini-2.5-pro	147	3.1%	Rows 2,496–2,640 (third parallel pass)
deepseek-v4-flash	~2,101	~44.4%	Rows 476– (labels in progress, Hermes Agent)
Empty (no label)	~0	~0%	Will be retried during background labeling run

Total re-labeled: 2,633 rows (55.6%) (+ ~2,101 deepseek in progress)

Label format

All three labelers were given the same prompt (the source text from stage0 chunks, truncated to 1,200 characters) and asked to produce a ≤5 word tag describing what concept the activation vector most likely encodes at the final token position.

Examples:

• opus-4-6: "Galileo refuting Aristotle's gravity"
• sonnet-4-6: "police body camera accountability"
• gemini-2.5-pro: "Kidney transplant candidate eligibility"

The labeler_model column in the parquet records which model generated each label, enabling downstream A/B analysis of labeler quality.

Schema

Same as stage3_v0_1_full/av_sft.parquet with one additional column:

Column	Type	Description
prompt	list[dict]	AV prompt template (unchanged)
response	string	Short label wrapped in <explanation>...</explanation>
labeler_model	string	One of: opus-4-6, sonnet-4-6, gemini-2.5-pro, gemini-2.5-flash-original
activation_vector	binary	Same as original
doc_id	string	Same as original

Known issues and fixes applied

Fixed

1. Corrupted chunk row_001798.json — truncated extra -4-6"} tail appended to valid JSON. Repaired 2026-05-15.
2. Missing labeler_model field on 862 Opus chunks — backfilled to opus-4-6 in both the chunk files and rebuilt parquet.

Unresolved

3. 261 Opus failures (opus_failed): 66 unique arxiv doc_ids, each with 4 consecutive failed rows. Likely Opus timing out on longer arxiv source texts.
4. 28 Gemini failures (gemini_failed): 7 kidney-donor doc_ids at the tail of Gemini's run where it hit QUOTA_EXHAUSTED.
5. 2 Sonnet failures (claude_failed): Negligible.

To fill the 291 gaps: ~36 minutes of Sonnet calls at 7.5s/row. The chunk directory is restart-safe — re-running the script skips completed rows.

Chunk directory

Raw per-row JSON files live in chunks/ (row_{index:06d}.json), each containing:

{
  "idx": 0,
  "doc_id": "fwe_00000053",
  "label": "Galileo refuting Aristotle's gravity",
  "labeler_model": "opus-4-6"
}

Chunks with empty labels have an "error" field instead.

Rebuilding the parquet

python experiments/v8_nla_local/stage2_opus_short_labels.py --n-rows 4734

The script reads chunks/ and rebuilds av_sft.parquet. It will also resume labeling any unfilled rows when run with no --n-rows limit.