---
license: mit
pretty_name: Scientific Post-Training DReaMS Evaluation v1
tags:
  - mass-spectrometry
  - molecule-retrieval
  - post-training
  - evaluation
---

# Scientific Post-Training DReaMS Evaluation v1

This repository contains the immutable public evaluation bundle used by the
`dreams-retrieval` Harbor task in ScientificPostTrainBench. It is evaluation
data, not agent training data.

## Files

- `public_spectra.hdf5`: three positive-mode `[M+H]+` MS/MS queries for each of
  128 target molecules.
- `public_candidates.jsonl.gz`: 256 same-formula candidates and target position
  for each molecule.
- `manifest.json`: upstream revision, source hashes, deterministic selection,
  and release-file hashes.

The bundle is deterministically derived from the MassSpecGym 1.5 test fold at
commit `c9aa3feb5f6ec0adee56cc78d2dce24826356156`. The source files and this
derived bundle are MIT-licensed. The builder is retained in
`ScientificPostTrainBench/authoring/dreams`.

Agents may query the bundle only through the task's `/app/evaluate.py` and read
aggregate metrics. Opening rows, extracting candidate targets, or fitting to
this bundle invalidates the benchmark trajectory.