MACE-Bench / REPRODUCE.md
tirtho149's picture
MACE-Bench release: images, memory banks, captions, predictions, traces, code
c8e03af verified
|
Raw History Blame Contribute Delete
2.95 kB

Reproducing MACE from this repository

Everything below runs from a copy of this repository. Steps 1 and 2 make no model calls and cost nothing.

1. Check every reported number (under a minute; numpy and pandas only)

python results/verify.py

Expected equal-crop accuracies (%), Sonnet 5.5, 1,600 queries over 10 crops:

Method In distribution Held-out collection
SAGE (KB only) 42.8 50.7
DINOv2 1-NN 52.5 38.2
Labels only 55.8 54.8
MACE retrieval 62.2 57.0

python results/verify_extension.py does the same for the follow-up tests in results/extension/.

2. Rebuild the memory banks and replay the model calls (no cost)

pip install -r code/requirements.txt
cd code
mkdir -p .cache && tar -xzf ../cache/llm.tar.gz -C .cache
AGMEM_DATA=$PWD/../benchmark python verify_release.py

This rebuilds all 12,438 evaluation jobs from benchmark/splits/ (memory selections in memory.parquet, folds in iid_5fold_3seed.parquet) and study/queries.csv, compares each job and its exact memory list with study/evaluation_jobs.jsonl.gz, then rebuilds every caption and text-diagnosis request and checks that the cached reply exists and gives the saved prediction. Expected last line: RELEASE REPRODUCES.

To recompute results/results.json from the raw predictions:

cp -R ../runs/P* ../captions.json ../results/local_baselines.csv mace10/out/
mkdir -p runs
# the ledger also holds later follow-up calls; the paper's price and latency rows use calls up to the analysis time
gunzip -c ../results/cost_ledger.jsonl.gz | python -c "import sys,json;[sys.stdout.write(l) for l in sys.stdin if json.loads(l)['t']<=1793870828.0]" > runs/cost_ledger.jsonl
AGMEM_DATA=$PWD/../benchmark python -m mace10.results

The new code/mace10/out/results.json equals results/results.json in all 1,062 values.

3. Rerun a package

With the cache from step 2 in place, any package replays from the cache without new calls:

cd code
AGMEM_DATA=$PWD/../benchmark python -m mace10.run P1 --cap-additional 0

Packages: P7dev P1 P4labels P2 P3 P4rest P5 P6 P7 P8 P9. Calls that miss the cache go to the Claude Code CLI (claude-sonnet-5-5); delete code/.cache/llm to rerun fresh, which costs about USD 185 at list prices. The agentic package P6 also needs tar -xzf ../cache/agentic.tar.gz -C .cache.

Memory banks

study/memory_banks.jsonl lists every bank as used: package, condition, crop, selection seed (0 to 2), fold run, K (1, 4 or 16 images per disease), the member image_ids with labels, and the queries it served. A query never sits in its own bank. The source of these banks is benchmark/splits/memory.parquet (facility-location selection, nested over K) filtered by the query's fold in benchmark/splits/iid_5fold_3seed.parquet; held-out collections are never in memory (benchmark/splits/ood_heldout_collection.parquet).