CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

A benchmark for evaluating vision-capable LLMs on Indian competitive exam questions (JEE Main, JEE Advanced, NEET). Questions are images sent to models via the OpenRouter API; responses are parsed from <answer>...</answer> tags and scored using exam-specific marking schemes.

Running the Benchmark

# Setup
uv sync
echo "OPENROUTER_API_KEY=your_key" > .env

# Must run from project root (paths are resolved relative to cwd)
uv run python src/benchmark_runner.py --model "google/gemini-2.5-pro-preview-03-25" --exam_name JEE_ADVANCED --exam_year 2025

# Filter by question IDs
uv run python src/benchmark_runner.py --model "openai/o3" --question_ids "N24T3001,N24T3002"

CLI args: --model (required), --exam_name (all/NEET/JEE_ADVANCED/JEE_MAIN), --exam_year (all/2024/2025), --question_ids, --output_dir, --config, --resume, --temperature (override config), --num_runs (default 1, use 3+ for variance analysis).

Analysis Scripts

# Generate cross-model leaderboard from all results
uv run python scripts/generate_leaderboard.py

# Aggregate multiple runs of the same model for variance analysis
uv run python scripts/aggregate_runs.py --pattern "openai_o3_JEE_ADVANCED_2025"

Testing

# Run the full pytest suite (67 tests)
uv run python -m pytest tests/ -v

# Run individual module self-tests
uv run python src/utils.py        # answer parsing logic
uv run python src/evaluation.py   # scoring logic
uv run python src/llm_interface.py  # API calls (requires .env and network)

Architecture

benchmark_runner.py  ─── orchestrator / entry point
  ├── loads config from configs/benchmark_config.yaml
  ├── loads dataset directly from metadata.jsonl (JSONL → HuggingFace Dataset)
  │     ├── metadata.jsonl (question metadata, 578 questions)
  │     └── images/ (question PNGs, stored in Git LFS)
  ├── calls llm_interface.py for each question
  │     ├── prompts.py (prompt templates)
  │     └── utils.py (parse_llm_answer extracts from <answer> tags)
  ├── scores via evaluation.py (exam-specific marking schemes)
  └── writes results incrementally to results/{model}_{exam}_{year}_{timestamp}/
        ├── predictions.jsonl  (raw API responses)
        ├── summary.jsonl      (scored per-question results)
        └── summary.md         (human-readable report)

Key data flow

1. Dataset loaded directly from metadata.jsonl into a HuggingFace Dataset object, filtered by exam/year
2. Each question image is base64-encoded and sent to OpenRouter with a structured prompt
3. If the response can't be parsed, a re-prompt is sent (text-only, with the bad response)
4. If the API call fails, the question is queued for retry (up to 3 attempts, exponential backoff via tenacity)
5. Answers are parsed from <answer>...</answer> tags by utils.parse_llm_answer()
6. evaluation.py scores using JEE/NEET marking schemes (partial credit for MCQ_MULTIPLE_CORRECT in JEE Advanced)

Answer Format Conventions

• MCQ_SINGLE_CORRECT: <answer>A</answer> → ["A"]
• MCQ_MULTIPLE_CORRECT: <answer>A,C</answer> → ["A", "C"] (sorted, deduplicated)
• INTEGER: <answer>42</answer> → ["42"]
• SKIP: <answer>SKIP</answer> → no penalty

Important Notes

• Git LFS: Images and metadata.jsonl are in LFS. Run git lfs pull after cloning.
• Working directory: Scripts must be run from project root — config, data, and image paths are resolved relative to cwd.
• Python 3.10+: Uses union type syntax (list[str] | str | None).
• Models: Configured in configs/benchmark_config.yaml under openrouter_models. All must support vision input.
• Result directory naming: results/{provider}_{model}_{exam}_{year}_{YYYYMMDD_HHMMSS}/ (slashes in model IDs replaced with underscores).