leetcrypt
58d405c518
Capture native-harness benchmark baselines for qwen2.5 0.5b/1.5b/3b and qwen2.5-coder:7b (all probed tool-capable; deepseek-r1 and NL2SH reject the tools field). All cluster at 1-2/12 with high variance; the 7B buys no pass-rate gain at ~3x latency, so qwen2.5:3b stays the default. The single biggest score sink across every model is bare tool-call-as-text leaks — a harness parse gap, the clear next improvement. Also drop "in your home directory" from the shell prompts: it made literal- minded models create a home/ subdir (/root/home/a/b/c/...) or use ~/, which the benchmark itself surfaced. Findings doc + bench README carry the model comparison table and recommendations (llama3.2:3b / llama3.1:8b for a non-qwen data point). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
86 lines
3.7 KiB
Markdown
86 lines
3.7 KiB
Markdown
# Native-harness benchmark
|
||
|
||
A small, ground-truth-graded benchmark for the `/ai <model> !<task>` native
|
||
tool-calling loop (`cmd_chat/agent/bridge.py:_run_native`). Use it to compare
|
||
harness changes (and models) **systematically** instead of eyeballing one-off
|
||
runs.
|
||
|
||
## Why ground truth
|
||
|
||
The weak CPU model routinely *claims* success it didn't achieve ("written, made
|
||
executable, and run successfully" after a 126; "created notes.txt" while the
|
||
file actually contains `echo 'hello'`). So every task is graded by a POSIX-sh
|
||
`verify` snippet run **inside the sandbox** via `podman exec`, exit 0 == PASS —
|
||
never by the model's own summary. See
|
||
`docs/findings-native-harness-2026-06-10.md`.
|
||
|
||
Tasks run in the agent's real cwd (the container HOME, `/root`) with **bare
|
||
filenames** — pinning an absolute `/root/bench/...` path instead just measures
|
||
whether the weak model honours absolute paths (it ~never does for `run_shell`
|
||
redirects) and drowns out every other signal. Only the suite's known artifacts
|
||
are wiped between tasks; the home dir itself is never `rm -rf`d.
|
||
|
||
## The matrix
|
||
|
||
Four categories × easy / medium / hard (`bench/tasks.py`):
|
||
|
||
| category | tools exercised | easy → hard |
|
||
|---|---|---|
|
||
| `shell` | run_shell | whoami→file · nested mkdir · count `/etc/*.conf` |
|
||
| `code` | write_file + run_shell | py `2+3` · **chmod +x script** · 10× Fibonacci |
|
||
| `git` | run_shell + network | clone · clone+read README · clone+report branch |
|
||
| `multi` | chained steps / recovery | mkdir→write→list · reverse-sort · word-count script |
|
||
|
||
`code-medium` (write → `chmod +x` → `./run`) and `multi-easy` are the
|
||
historically hardest cases — they're the nudge loop's main targets, so they
|
||
double as regression canaries.
|
||
|
||
## Prerequisites
|
||
|
||
Same setup as a manual live test:
|
||
|
||
1. Server + Rust client running in a tmux window (default `hh-house:2`).
|
||
2. An agent online with drive: `/ai start <model>` then `/grant ai`.
|
||
3. The sandbox container running (default `hack-house`) with `python3`, `git`,
|
||
and outbound network (the `git` tier needs it).
|
||
|
||
The runner only *sends* tasks and *reads* state — it never starts/stops the
|
||
agent, so a failure leaves your session intact.
|
||
|
||
## Running
|
||
|
||
```bash
|
||
.venv/bin/python bench/run.py --model qwen2.5:3b # full suite
|
||
.venv/bin/python bench/run.py --tier easy # one tier
|
||
.venv/bin/python bench/run.py --no-net # skip git/network tasks
|
||
.venv/bin/python bench/run.py --only code-medium,multi-easy
|
||
.venv/bin/python bench/run.py --dry-run # print the matrix, run nothing
|
||
```
|
||
|
||
Useful flags: `--container`, `--target` (tmux window), `--user` (for @mention
|
||
detection), `--timeout` (per-task wallclock cap, default 300 s).
|
||
|
||
## Output
|
||
|
||
A PASS/FAIL table with per-task latency, plus a JSON record under
|
||
`bench/results/` (gitignored — these are runtime artifacts). Commit a curated
|
||
baseline explicitly (e.g. `--out bench/results/baseline-<model>.json` then
|
||
`git add -f`) when you want one tracked for comparison.
|
||
|
||
## Tracked baselines (2026-06-10, CPU-only, shared room)
|
||
|
||
| model | size | PASS | avg/med s |
|
||
|---|---|---|---|
|
||
| qwen2.5:0.5b | 397 MB | 1/12 | 55 / 49 |
|
||
| qwen2.5-coder:1.5b | 986 MB | 1/12 | 67 / 53 |
|
||
| **qwen2.5:3b** | 1.9 GB | 1/12 | 51 / 46 |
|
||
| qwen2.5-coder:7b | 4.7 GB | 2/12 | 141 / 106 |
|
||
|
||
All cluster at 1–2/12 with high run-to-run variance — treat as a *relative*
|
||
regression yardstick, not an absolute grade. The 7B buys no pass-rate gain at
|
||
~3× latency; **qwen2.5:3b stays the default**. The biggest score sink across
|
||
every model is bare tool-call-as-text leaks (a harness parse gap, not model
|
||
capability) — see `docs/findings-native-harness-2026-06-10.md`. Only
|
||
qwen2.5(-coder) sizes 0.5b–7b are usable here; `deepseek-r1` and
|
||
`westenfelder/NL2SH` reject Ollama's `tools` field.
|