ducoterra
f221b40fce
New files: - scripts/model_benchmark.py — shared CSV recorder for all model tests - scripts/test_summary_model.py — summary model quality benchmark (coherence, coverage, brevity, hallucination) - scripts/test_embed_model.py — embedding model benchmark (dimension, cosine accuracy, speed) - .agents/skills/test-summary-model/SKILL.md — skill for testing summary models - .agents/skills/test-embed-model/SKILL.md — skill for testing embedding models - benchmarks/README.md — schema documentation Updated: - .agents/skills/test-chat-model/SKILL.md — now also records to CSV All three scripts write to benchmarks/model_benchmarks.csv with one row per run per check. The CSV accumulates results across runs for comparison.
100 lines
3.6 KiB
Markdown
100 lines
3.6 KiB
Markdown
---
|
||
name: test-summary-model
|
||
description: Tests a summary model (BOR_LLM_SUMMARY_MODEL in .env) against a fixed set of 8 source texts from the fixture KB — evaluates coherence, coverage, brevity, and hallucination detection — then records results in benchmarks/model_benchmarks.csv. Use when the user asks to test or benchmark a summary model, compare summary models, or evaluate summary quality (e.g. "test lite summary", "compare summary models", "how good is turbo at summarizing").
|
||
---
|
||
|
||
# Test a Summary Model (quality benchmark)
|
||
|
||
Tests the configured `BOR_LLM_SUMMARY_MODEL` (default `lite`) against
|
||
8 source texts extracted from the fixture KB. Each text is a realistic
|
||
documentation excerpt (~300–500 chars). The model is asked to summarize
|
||
each in 2–4 sentences.
|
||
|
||
## Rules (non-negotiable)
|
||
|
||
- **Do not touch** the fixture source texts — they are the controlled
|
||
benchmark corpus. Changing them is a methodology change: flag it.
|
||
- **Do not edit app code** (`app/`, `tests/`). This skill tests a model,
|
||
not the app. If the model exposes an app defect, report it — don't fix it.
|
||
- `.env` is gitignored and is **not committed** — the model switch stays
|
||
a live dev setting, and the summary must say which model `.env` is left
|
||
on (default: the tested model).
|
||
|
||
## Procedure
|
||
|
||
All commands run from the repo root with `uv run`.
|
||
|
||
### 1. Preconditions
|
||
|
||
```bash
|
||
podman compose up -d db
|
||
grep -E "BOR_LLM_(BASE_URL|API_KEY|SUMMARY_MODEL)" .env
|
||
```
|
||
|
||
### 2. Switch the model (optional)
|
||
|
||
Edit only `BOR_LLM_SUMMARY_MODEL` in `.env` (leave chat and embed alone):
|
||
|
||
```
|
||
BOR_LLM_SUMMARY_MODEL=<model>
|
||
```
|
||
|
||
### 3. Run the benchmark
|
||
|
||
```bash
|
||
# Single run
|
||
uv run python -m scripts.test_summary_model
|
||
|
||
# Specify a model explicitly
|
||
uv run python -m scripts.test_summary_model --model turbo
|
||
|
||
# Multiple runs for variance
|
||
uv run python -m scripts.test_summary_model --runs 3
|
||
```
|
||
|
||
Wall time: ~3–5 s per text, ~25–40 s total for 8 texts.
|
||
|
||
### 4. Read the verdicts
|
||
|
||
The `gate:` line: `PASS|FAIL turns=N answered=N quality=Q hallucinations=H/N coherence=C/5 coverage=V/5 brevity=B/5 (wall Ts)`.
|
||
|
||
Scoring rubric:
|
||
- **coherence** (0–5): non-empty, multi-sentence, starts with capital
|
||
- **coverage** (0–5): captures ≥2 key numbers/facts from source
|
||
- **brevity** (0–5): summary length / source length ratio 0.10–0.25 = 5, 0.05–0.35 = 4, etc.
|
||
- **hallucination**: detected when summary contains >3 uncommon tokens not in source
|
||
|
||
Quality score = coherence×0.35 + coverage×0.40 + brevity×0.25, scaled 0–100,
|
||
with a 5-point penalty per hallucination.
|
||
|
||
Gate: PASS if quality ≥ 70 and hallucination rate < 25%.
|
||
|
||
### 5. Record + commit
|
||
|
||
Results are automatically appended to `benchmarks/model_benchmarks.csv`.
|
||
Also append a summary line to `benchmarks/README.md` if it exists, or
|
||
create it:
|
||
|
||
```bash
|
||
git add benchmarks/model_benchmarks.csv
|
||
git commit --no-gpg-sign \
|
||
-m "docs(agent): record the <model> summary benchmark" \
|
||
-m "<one-paragraph body: quality score, hallucination count, wall time, comparison to other models>"
|
||
```
|
||
|
||
### 6. Summary
|
||
|
||
Table of quality / hallucinations / coherence / coverage / brevity / wall
|
||
per run, the one-line conclusion, and a note that `.env` is now on
|
||
`<model>`.
|
||
|
||
## Troubleshooting
|
||
|
||
- **endpoint down / all turns error** — check
|
||
`curl -s $BOR_LLM_BASE_URL/models` with the key; the gate will exit 1
|
||
with `answered<8`. Report, don't retry-loop.
|
||
- **low brevity** — the model is restating rather than condensing.
|
||
This is model-specific; try a more explicit prompt or a different model.
|
||
- **hallucinations** — the model is adding details not in the source.
|
||
This is the most common failure mode; tighten the prompt or switch models.
|