feat: Add extract_eval_stats.py script for eval statistics extraction

[geelen]

Summary

I found myself wanting to understand the performance and caching behavior of a benchmark run a bit better. So I had this script made. Right now it's just a standalone script, but it could be integrated into the main app if people found it useful.

What are you adding?

• Bug fix (non-breaking change which fixes an issue)
• New benchmark/evaluation
• New model provider
• CLI enhancement
• Performance improvement
• Documentation update
• API/SDK feature
• Integration (CI/CD, tools)
• Export/import functionality
• Code refactoring
• Breaking change
• Other

Changes Made

New scripts/extract_eval_stats.py standalone script that pulls info from the log file.

Testing

• I have run the existing test suite (pytest)
• I have added tests for my changes
• I have tested with multiple model providers (if applicable)
• I have run pre-commit hooks (pre-commit run --all-files)

Checklist

• My code follows the project's style guidelines
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation (if applicable)
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

---

Extract detailed statistics from openbench/Inspect AI evaluation logs (.eval ZIP or .json files).

• Total time and working time per sample

• LLM HTTP round-trip time (from model call events)

• Provider-specific timing for Groq (queue_time, prompt_time, completion_time)

• Approximate tool time (working_time - llm_http_time)

• LLM time fraction of total working time

• Input, output, and total tokens per sample and aggregate

• Per-model breakdown from model_usage

• OpenAI: prompt_tokens and cached_tokens with cache hit rate

• Anthropic: cache_read_input_tokens and cache_creation_input_tokens

• Human-readable summary to stdout

• JSON output with --output flag

• Per-sample statistics with --per-sample flag

python scripts/extract_eval_stats.py logs/my_eval.eval

python scripts/extract_eval_stats.py logs/ -o stats.json --per-sample -v

  Model: groq/openai/gpt-oss-120b
  Task: progressivemcpbench-copilot

  Samples: 130
  Score: mean=0.523 (min=0.000, max=1.000)

  CALLS:
    Model calls: 794
    Tool calls: 590

  TIME BREAKDOWN:
    Total eval time: 1h 46m
    LLM time: 7m 49s (7.3%)
    Tool time (approx): 1h 38m

  LLM CALL PERFORMANCE (avg per call):
    HTTP round-trip: 623ms
    Time to first token (queue): 12ms
    Prompt processing: 55ms
    Completion generation: 307ms

  TOKENS:
    Input: 5.41M
    Output: 124.4K
    Total: 5.54M

  THROUGHPUT:
    Input processing: 74.2K tok/s
    Output generation: 475 tok/s

  CACHING (OpenAI-compatible):
    Prompt tokens: 5.41M
    Cached tokens: 3.94M
    Cache hit rate: 72.8%

• Excludes scorer model calls (only counts primary model usage)
• All data is extracted from existing eval logs - no runtime changes required
• Cache statistics depend on provider support (OpenAI, Anthropic formats)

Amp-Thread-ID: https://ampcode.com/threads/T-2f80b096-55b1-4353-930c-76e9f6368f71

---

Note

Introduce a standalone script to parse .eval/.json logs and output aggregated/per-sample eval timing, token, throughput, and cache statistics.

• CLI (scripts/extract_eval_stats.py):
  • Parse .eval (ZIP) and .json logs; auto-discover files in a path.
  • Extract per-sample stats (times, token usage), excluding scorer model calls via span hierarchy.
  • Aggregate eval metrics: total/working time, LLM HTTP time, tool-time approximation, call counts.
  • Compute provider timings (Groq: queue_time, prompt_time, completion_time) and throughput (input/output tok/s).
  • Collect cache metrics: OpenAI (prompt_tokens, cached_tokens, hit rate) and Anthropic (cache_read_input_tokens, cache_creation_input_tokens).
  • Output human-readable summary to stdout; optional JSON via --output with --per-sample support; verbose mode.

^{Written by Cursor Bugbot for commit ce075e0. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Agent can help with this pull request. Just @cursor in comments and I'll start working on changes in this branch.
_{Learn more about Cursor Agents}

[geelen]

@cursor please fix both suggestions.

[cursor]

Provider timing averages now weight every individual call rather than averaging per-sample means, so llm_queue/prompt/completion/provider_time_mean mirror the llm_http_time_mean behavior and avoid bias toward short samples.

        all_http_times.extend(s.llm_http_times)
        if s.n_model_calls > 0:
            total_queue_time_sum += s.llm_queue_time_total
            total_prompt_time_sum += s.llm_prompt_time_total
            total_completion_time_sum += s.llm_completion_time_total
            total_provider_time_sum += s.llm_provider_time_total
            total_provider_time_calls += s.n_model_calls
...
            "llm_queue_time_mean": (
                total_queue_time_sum / total_provider_time_calls
                if total_provider_time_calls > 0
                else None
            ),

The P95 calculation now uses an interpolated percentile index (0.95 * (n-1)) so 95th percentile isn’t biased high for small samples.

        total_time_p95 = None
        if len(total_times) >= 20:
            sorted_times = sorted(total_times)
            percentile_position = 0.95 * (len(sorted_times) - 1)
            lower_index = int(percentile_position)
            upper_index = min(lower_index + 1, len(sorted_times) - 1)
            fraction = percentile_position - lower_index
            total_time_p95 = (
                sorted_times[lower_index] * (1 - fraction)
                + sorted_times[upper_index] * fraction
            )
...
            "total_time_p95": total_time_p95,

Let me know if you’d like this scripted into tests or covered elsewhere.

---

 

_{Learn more about Cursor Agents}