generality-labs/glm_features

Per-transcript GLM features

Deterministic features of every transcript in LASR-G5/benchmarks, for the transcript-regression protocols (glm_length, glm_regex) in small-judge. Produced by scripts/glm_features/extract.py from the benchmark logs at commit c199498b; no generative model is involved.

Layout

Mirrors the benchmarks repo, one CSV per successful full-run log:

<benchmark>/<task_args_hash>/<model>/<repeat>/<eval id>.csv

Crashed or partial logs kept beside a successful run are not extracted. Only the benchmarks used by the four domains are present: agent_bench_os, swe_bench_mini, terminal_bench_2_1 (coding); cybench, intercode (cyber); aime2024/2025/2026, apex_2025, apex_shortlist, hle, mmlu_pro_math (math); arc_challenge, gpqa, hle (science).

Rows and columns

Two rows per sample: reasoning = visible (assistant text plus raw chain of thought) and hidden (assistant text only). Selection of items, epochs and models, and averaging per model, happen downstream.

Column	Meaning
item, epoch, correct	sample id, epoch, and whether the first scorer marked it correct
log_chars	ln(1 + characters of assistant text, including visible reasoning)
log_reasoning_chars, visible_reasoning	ln(1 + reasoning characters); 1 if any raw reasoning is visible
personal_pronouns, hedging, backtracking, uncertainty, confidence, questions, verification, connectives, contrast	phrase hits per 1,000 prose words, after removing code, maths and ANSWER: lines
log_tool_calls	ln(1 + structured tool calls); NaN when the log has no message history

A sample with no assistant message has NaN in every feature. Definitions and term lists: docs/transcript-feature-table.md and src/protocols/transcript_features/features.py in small-judge.