Claude/rlm testing framework byf lb #14
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add an extensible framework to recreate tests from the original RLM paper (arXiv:2512.24601) and support new benchmark development. Framework components: - benchmarks/base.py: Abstract Benchmark class, BenchmarkSample, BenchmarkResult - benchmarks/metrics.py: Evaluation metrics (exact match, containment, token F1, pairwise F1) - benchmarks/runner.py: BenchmarkRunner with support for RLM, direct LLM, and summarization methods - benchmarks/cli.py: CLI for running benchmarks and comparing methods Benchmark implementations: - NIAHBenchmark: Single-Needle-in-a-Haystack (synthetic retrieval) - OolongBenchmark: Semantic aggregation using HuggingFace oolongbench dataset - OolongPairsBenchmark: Pairwise combinatorial aggregation (hardest setting) - BrowseCompPlusBenchmark: Multi-hop QA over document corpora (synthetic) Usage: python -m benchmarks.cli --benchmark oolong --methods rlm direct -n 10 python -m benchmarks.cli --benchmark all --output results.json
Add a results storage system that enables: - Persistent storage of benchmark results in JSON-lines format - Historical comparison across experiment runs - Query by benchmark, model, method, or environment - CSV export for external analysis tools New CLI subcommands: - `run`: Run benchmarks (default, backward compatible) - `history`: Show historical results with filters - `compare`: Compare results grouped by method/model/environment - `list`: Show summary of stored results - `export`: Export benchmark results to CSV Results are automatically saved to ./benchmark_results/ directory with full experiment metadata including git commit and RLM version.
Add --max-workers (-w) option to run samples concurrently using ThreadPoolExecutor. This significantly speeds up evaluation when making many independent LLM API calls. Usage: # Run 4 samples in parallel python -m benchmarks.cli run --benchmark niah -n 20 --max-workers 4 # Or via Python API runner = BenchmarkRunner(backend="openai", model="gpt-5", max_workers=4) results = runner.run(benchmark, num_samples=100) Default is sequential (max_workers=1) for predictable behavior.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.