Model Accuracy Benchmark

Model Accuracy Benchmark

Run an accuracy (eval) pass on a newly adapted model after the serving smoke test has passed. By default run on the user-specified TPU machine or remote environment, not on the local development checkout, unless the user explicitly asks for a local-only dry run.

This skill covers accuracy only. Throughput/latency benchmarks belong to model-speed-benchmark. The two are independent contracts: different triggers, required inputs, validation checks, and output fields. If the user asks to "run both", invoke the two skills explicitly and produce one report each.

Workflow

1. Confirm the target resource, repo path/source, branch/commit, model path, host, and port.
2. Confirm the model server is already running on the target device and verify it is live. /health returns HTTP 200 with an empty body, so check the status code — not the body — (e.g. curl -s -o /dev/null -w '%{http_code}' http://HOST:PORT/health), or treat a populated /v1/models response as the liveness signal.
3. Record the first served model id from /v1/models and check whether another eval is already running against the same endpoint.
4. Collect the serving process id (PID) that owns the endpoint port, so Server id in the report is never left as unknown when the process is reachable. See "Collecting the server id".
5. Start a fresh, unique work directory for the current request, following "Output directory layout".
6. Run the accuracy benchmark with evalscope eval (preferred path).
7. If evalscope cannot be installed or invoked on the target environment, fall back to scripts/accuracy_benchmark.sh (wraps test/srt/run_eval.py).
8. If the target machine does not have the wrapper scripts, copy scripts/ to the target checkout first.
9. Validate the produced artifacts before reporting them.
10. Write the final markdown report by filling templates/report-template.md (see "Output format").

Collecting the server id

Server id is a required report field. Do not write unknown unless every probe below fails. /health and /v1/models do not expose a PID, so query the process that owns the serving port on the target host:

ssh <host> 'bash -lc "ss -ltnp 2>/dev/null | grep :<port> || sudo ss -ltnp 2>/dev/null | grep :<port>; pgrep -af launch_server | head"'

Report the PID of the process listening on the endpoint port (for the sgl-jax launcher this is the sgl_jax.launch_server process). Use exp_id instead of PID only when the run is managed by an experiment runner that exposes one. Write unknown only when the process is genuinely unreachable, and say so in ## Notes.

Output directory layout

Use one timestamped, unique directory per request. For a standalone accuracy run a flat directory is fine:

/tmp/evalscope_<dataset>_<limit-or-full>_<YYYYMMDD_HHMMSS>/

Generate the timestamp on the target host (for example date +%Y%m%d_%H%M%S). Keep --work-dir unique per run/per report to avoid mixing artifacts across runs.

Accuracy Benchmark

Preferred path: run evalscope eval directly on the target host. Use this path whenever evalscope is available via uv tool install evalscope or is already installed. This is the default for all accuracy benchmarks, including datasets not wired into the repo-local run_eval.py wrapper.

Default command:

evalscope eval --model <served model id or path> --api-url http://HOST:PORT/v1/chat/completions --eval-type openai_api --datasets <dataset> --eval-batch-size 16 --generation-config '{"max_tokens":4096,"temperature":0.0}' --work-dir <unique work dir> --no-timestamp

Defaults:

Examples:

evalscope eval --model /models/mistralai/Mistral-7B-Instruct-v0.3 --api-url http://127.0.0.1:30000/v1/chat/completions --eval-type openai_api --datasets aime24 --limit 30 --eval-batch-size 16 --generation-config '{"max_tokens":4096,"temperature":0.0}' --work-dir /tmp/evalscope_aime24 --no-timestamp
evalscope eval --datasets gsm8k --limit 10 --work-dir /tmp/evalscope_gsm8k_10 --no-timestamp

Compatibility and quoting notes:

• The raw JSON string form --generation-config '{"max_tokens":4096,"temperature":0.0}' is confirmed working on evalscope 1.8.1. Prefer it.
• Some older evalscope versions expect the key=value,key=value form (--generation-config max_tokens=4096,temperature=0.0). If the JSON form is rejected, probe with evalscope eval --help or a one-sample run, then switch formats explicitly.
• SSH quoting hazard: nesting ssh <host> 'bash -lc "... --generation-config {...} ..."' strips the inner JSON quotes and breaks the argument (ValueError: not enough values to unpack). Write the command to a small remote script and run that, or otherwise guarantee the JSON quotes survive both shell layers.
• Treat evalscope CLI incompatibility as an invocation failure. If it cannot be resolved quickly, use scripts/accuracy_benchmark.sh and note the deviation in ## Notes.

Use concrete dataset names in --datasets, for example gsm8k, aime24, or aime25. Prefer --api-url for normal OpenAI-compatible endpoints. Pass --model when the served model id should be explicit in the report or when auto-discovery is unreliable.

Fallback path: use scripts/accuracy_benchmark.sh only if evalscope cannot be installed or invoked on the target host. This wrapper calls test/srt/run_eval.py (not evalscope eval) and supports only the benchmark names exposed by that repo-local runner. Its artifact layout differs from evalscope: run_eval.py writes /tmp/<eval_name>_<model>.json and /tmp/<eval_name>_<model>.html. When using the fallback, validate and report those /tmp files instead of the evalscope reports/<model_id>/<dataset>.json path.

Fallback command:

scripts/accuracy_benchmark.sh --repo-dir <sglang-jax root> --python-bin <repo>/.venv/bin/python --base-url http://127.0.0.1:30000 --eval-name gsm8k --num-examples 30 --max-tokens 4096 --temperature 0.0

Existing artifacts

Do not reuse old eval artifacts as the result for a new request unless the user explicitly asks for historical comparison or artifact inspection. For normal requests: always run a fresh eval; always use a new unique work directory; do not report an old artifact as if it were produced by the current run. Old artifacts may still be inspected for debugging, validation, or comparison, but they do not replace the current run.

Artifact validation

Before reporting success, validate the outputs that were actually produced:

• The work directory exists.
• For evalscope, reports/<model_id>/<dataset>.json exists and is readable.
• For the fallback wrapper, /tmp/<eval_name>_<model>.json and its matching .html report exist and are readable.
• The reported sample count matches the requested --limit when a limit was specified.
• The score can be read from the report artifact.

If validation fails after a run completed, report partial or failed instead of claiming success.

Output format

The report shape is defined by templates/report-template.md. Fill every placeholder and keep the section order, headings, and field labels exactly as in the template — it is the single source of truth for this skill's output. Write the report in English (headings, field labels, notes); any brief blocking/failure note emitted mid-run should also be in English.

By default, do not show incremental progress: run the full eval first, then present one consolidated report. Only interrupt this default when the run is blocked and needs user input, the run failed and the failure itself is the result, or the user explicitly asks for live progress.

Requirements for filling the template:

• Repeat one ### {dataset} block per dataset that was actually run.
• In the command block, include the concrete command that was actually executed.
• Do not omit fields within a section. If a field is unavailable, write unknown.
• In ## Status, always include target resource, endpoint, model path, server health, the known PID/exp_id, the Hardware group, the Versions group (engine (sglang-jax), evalscope, jax, libtpu), and the Server launch args.
• Hardware: fill TPU type (e.g. tpu-v6e), topology (e.g. 4x4), and chip count (e.g. 16). Take these from the target resource spec the user provided, or from the host's TPU metadata; write unknown for any that cannot be determined.
• engine (sglang-jax) is the served engine's version, git commit, and branch — get the commit with git -C <repo> rev-parse --short HEAD, the branch with git -C <repo> rev-parse --abbrev-ref HEAD, and the package version via <repo>/.venv/bin/python -c "import sgl_jax; print(sgl_jax.__version__)" when available; format as <version> @ <commit> (<branch>). This ties the score to a specific build for regression tracking.
• Server launch args: the full command line that started the server. Reuse the pgrep -af launch_server / ps -o args= -p <PID> output from "Collecting the server id" — the sgl_jax.launch_server command line is exactly the launch args, and its flags (--tp-size, --attention-backend, --dtype, --page-size, ...) can affect scores. Write unknown only if the process is unreachable.
• Prefer concrete version command results from the target environment; do not assume all versions come from the same environment.
• Collect evalscope as a uv tool install first, not from the project .venv. On remote hosts run version checks through a login shell (ssh <host> 'bash -lc "evalscope --version || uv tool list"') so ~/.local/bin is present; if still not found, explicitly try ~/.local/bin/evalscope --version. Only write unknown if all checks fail.
• Collect jax and libtpu from the remote project runtime (typically <repo>/.venv/bin/python) so they reflect the actual eval execution environment.
• Do not use .venv/bin/python -c 'import importlib.metadata ...' to infer the evalscope version — the uv tool install environment is separate from the project virtualenv and this lookup can be wrong.
• Prefer one concise Artifacts list with the primary result paths (the original paths on the target machine by default). Do not mention local copies unless the user explicitly asks.
• In ## Notes, explain any deviation from the full workflow, for example using a wrapper, inspecting an old artifact for debugging only, or a non-value-changing header cleanup.
• Unless the user explicitly asks for live progress, do not emit in-progress sections with unknown/pending/collecting placeholders before execution actually completes.

State selection rules:

• success: every requested dataset finished and its artifacts passed validation.
• partial: at least one dataset produced usable artifacts but another failed, was interrupted, or is missing validated artifacts.
• failed: no dataset produced usable, validated artifacts.

Retry: if accuracy fails before producing a valid report and the failure is clearly due to command construction or CLI argument format, retry once with a corrected invocation; prefer a fresh work directory for retries.

Remote setup

Sync the wrapper when it is missing on the target host (run from the repo root):

scp .claude/skills/model-accuracy-benchmark/scripts/accuracy_benchmark.sh <target_host>:<sglang-jax root>/scripts/

Install evalscope when it is missing on the target environment:

uv tool install evalscope