HealthAgentBench original run outputs
Lossless archives of the original run outputs associated with the HealthAgentBench leaderboard. Leaderboard: https://microsoft.github.io/HealthAgentBench/#leaderboard
The package contains 77,821 files (88,699,684,240 source bytes) in 40 independent tar.zst shards across 12 agent/model combinations. The shards total 50,763,152,396 bytes (50.76 GB), a 42.8% reduction. Each shard is at most 2 GiB.
Contents and provenance
Sources were downloaded from Azure account hanoverdev, container scratch, under qianchuliu/medcli/results/. The selection includes 70 category-specific run directories and the two shared roots opus-5/ and gpt-5.6-sol/, which contain eight dated run directories. Original relative paths are preserved inside the archives.
Archives include all downloaded artifacts, including failed attempts, retries, logs, and auxiliary files. The set of archived trials can therefore differ from the trials used to calculate leaderboard scores. This package does not recompute those scores.
Each trial directory is wholly contained in one shard. Shards are grouped by harness and model. Run-level metadata is assigned once and may be in a different shard from some trials in that run. Download all shards for a model to recover its complete run trees.
File contents, paths, modes, and modification times are retained. Archive ownership is normalized to uid/gid 0 with empty owner/group names for portable extraction. The source files were not edited or deleted.
This repository is an archive distribution. Its raw tar.zst shards do not provide a tabular Hugging Face dataset viewer.
manifest.json: models, shards, source directories, and trial/metadata assignments.files.jsonl: one record per source file, with archive path, byte size, SHA-256, and shard.SHA256SUMS: checksums for all shards and accompanying metadata files.verification.json: verification totals and results.shards/: independent archives compressed with Zstandard level 3.
Verify and extract
Install GNU tar and Zstandard. From the root of the downloaded package:
sha256sum -c SHA256SUMS
mkdir -p restored
for archive in shards/*.tar.zst; do
tar --zstd --no-same-owner -xf "$archive" -C restored
done
The restored tree is restored/healthagentbench-results/. Every shard is independently extractable; no concatenation is necessary. To extract only a model, use its filename prefix in the loop, for example shards/codex--gpt-5.5--*.tar.zst.
Upload
Replace NAMESPACE/REPO with your Hugging Face dataset repository. From the package root:
hf auth login
hf upload NAMESPACE/REPO . . --repo-type dataset
Verification
During packaging, each source file was hashed as its bytes entered the tar stream. Every completed archive was fully decompressed and its file names, sizes, and SHA-256 hashes were compared with those source hashes. Zstandard stream integrity also passed. All source files occur exactly once across the archive set. Parent directory headers can appear in multiple shards to support independent extraction.
Models
| Agent | Shards | Source GB | Compressed GB |
|---|---|---|---|
| Claude Code (Opus 5) | 5 | 11.26 | 2.63 |
| Codex (GPT-5.6-sol) | 7 | 15.91 | 12.28 |
| Codex (GPT 5.5) | 3 | 7.09 | 4.89 |
| Copilot (Opus 4.8) | 2 | 4.06 | 2.71 |
| Copilot (GPT 5.5) | 4 | 9.93 | 6.75 |
| Claude Code (Opus 4.8) | 2 | 2.83 | 0.63 |
| Codex (GPT 5.4) | 4 | 9.81 | 6.83 |
| Claude Code (Opus 4.7) | 2 | 3.79 | 0.76 |
| Codex (GPT 5.3) | 6 | 14.05 | 9.51 |
| Claude Code (Opus 4.6) | 2 | 3.48 | 0.72 |
| Claude Code (Sonnet 4.6) | 1 | 2.61 | 0.53 |
| Codex (GPT 5.4 Mini) | 2 | 3.88 | 2.53 |
- Downloads last month
- 18