File size: 5,508 Bytes
f8c91b8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | # TCB: Trust-Calibration Benchmark for Language-Model Workers
A small, auditable qualification test for language models used as components of multi-LLM
pipelines: 60 yes/no items across five domains (arithmetic, dates, geography, science,
language), each tagged easy or hard, with audited gold answers. A model answers
`ANSWER: Yes|No` and `CONF: 0-100`; we score accuracy, expected calibration error (ECE),
Brier, yes-rate, strict-format adherence, and availability, under an AUDITED administration
protocol (template-guarded parsing, answers only from tagged or bare final lines,
escalate-on-unparseable budgets, imputation flags).
## Files
- `../data/tcb_benchmark.json` — the 60 items (id, question, gold truth, tier), the prompt,
and scoring definitions. v0.2.1 corrects one gold label (g7: Astana IS the capital of
Kazakhstan), found by the independent audit of all 60 labels (`../data/label_audit.json`).
- `../data/tcb_v03_trials.json` — per-trial records, validated (adequate) protocol.
- `../data/tcb_v03_trials_constrained.json` — audited parser, fixed 512-token budget,
no escalation (unparseable = unavailable).
- `../data/tcb_v03_trials_naive.json` — the naive-protocol records. The naive and constrained
protocols consumed the SAME generations; `derive_naive.py` re-derives every naive verdict
offline from the constrained file's full raw responses using the historical parser
(preserved verbatim inside it), with 0/540 mismatches against the original run.
- `../data/tcb_repeat_results_naive.json` — the naive-protocol resampling summary (the
fabricated flip rates), alongside the validated-protocol repeat files.
- `../data/tcb_harness_comparison.json` — the three-protocol comparison summary.
- `../data/tcb_repeat_{results,trials}.json` — run-to-run variation (3 repeats, temp 0.7).
- `../data/tcb_decouple{,_trials}.json` — the confidence-elicitation manipulation.
- `../data/tcb_criterion{,_trials}.json` — the transfer study (SST-2 annotation + battery).
- `analyze_panel.py` — canonical analyzer: Table 1, figures, pairwise exact-McNemar + BH,
from the shipped validated-protocol trials. Stdlib + matplotlib only.
- `analyze_harness.py` — the three-protocol harness-validity comparison (figure + JSON).
- `analyze_decouple.py` — Q3 re-analyzer: per-model per-variant accuracy/ECE/mean-confidence
from `tcb_decouple_trials.json`, checked against `tcb_decouple.json` (stdlib only, no endpoint).
- `analyze_criterion.py` — Q4 re-analyzer: per-task accuracy/ECE, benchmark-to-downstream
Spearman with exact permutation p-values, gate split, and gate-pass/fail calibration ratios
with and without the leanstral-2603 outlier, from the shipped trials (stdlib only, no endpoint,
no HuggingFace download); Spearman values checked against `tcb_criterion.json`.
- `analyze_classwise.py` — camera-ready additions: per-model confusion matrices,
per-class precision/recall/F1, macro-F1, balanced accuracy (writes
`../tables/tcb_classwise.tex`, with pooled validated and naive rows), plus the
confidence-resolution statistics (accuracy at stated confidence below 100 vs at 100,
pooled AUROC of stated confidence for correctness) and the panel accuracy-vs-ECE
Spearman correlation. Stdlib only.
- `sensitivity.py` — ECE with vs without confidence-imputed trials.
- `tcb_parsing.py` — the audited parser + escalation rule (single canonical implementation).
- `run_tcb.py` — evaluate a new model (one command; appends to the panel).
- `run_panel_v03.py` (in the project tree), `run_repeat_v03.py`, `run_decouple.py`,
`run_criterion.py` — generation scripts (need an endpoint; all outputs ship).
- `analyze_tcb.py` — historical v0.2 analysis, kept for provenance; writes its table fragment
to `tables/tcb_results_v02.tex` so it never overwrites the canonical 9-model `tcb_results.tex`.
## Reproduce the paper (no model queries needed)
```bash
python3 analyze_panel.py # Table 1, reliability + forest figures, pairwise tests (Q1)
python3 analyze_harness.py # the three-protocol comparison (Figure 3 + summary JSON, Q2)
python3 analyze_classwise.py # confusion matrices + per-class P/R/F1 table, confidence resolution
python3 analyze_decouple.py # the confidence-elicitation manipulation (Q3)
python3 analyze_criterion.py # the transfer study: Spearman + permutation p + gate split (Q4)
python3 sensitivity.py # imputation sensitivity
python3 derive_naive.py # re-derive the naive condition from shipped raws (prints 0/540)
```
Every table, figure, and statistical test in the paper is reproduced offline by the commands
above from the shipped per-trial data; only the `run_*.py` generators need a model endpoint.
## Evaluate a new model / audit a new harness
```bash
python3 run_tcb.py <model-name> > my_trials.json
```
Append the rows to `tcb_v03_trials.json` and re-run `analyze_panel.py`; extra models are
detected automatically. To audit a different harness, run its parser over the shipped raw
responses and compare verdicts against `tcb_v03_results.json`.
## Metrics
- Accuracy (Wilson 95% CI), over parseable responses; availability reported separately.
- ECE, 5 equal-width bins, bootstrap CIs; imputed-confidence trials count (a missing
confidence is an unusable signal), with stated-only sensitivity via `sensitivity.py`.
- Brier with bootstrap CIs; pairwise exact binomial McNemar, Benjamini-Hochberg corrected.
- Yes-rate vs the 0.60 gold base rate; strict-format adherence out of 60.
|