Driver metadata
- Slug
- qwen25
- Status
- active
- Protocol match
- unknownDriver measures a metric the paper does not directly report. Validator C1 auto-downgrades not_reproduced to pending.
- Agent version
- v0.1.0-qwen25-winogrande-microslice
- Paper
- Qwen2.5 Technical Report
- Model
- Qwen/Qwen2.5-0.5B-Instruct
Claim citation
This is the exact paper claim the driver compares against before the Verdict Validator allows any public WRONG path.
- Location
- Results · Qwen2.5-0.5B-Instruct
- Metric
- MMLU 5-shot(accuracy)
- Reported value
- 47.5
- PDF page
- 8
- Quoted text
- “47.5”
Source
Every driver is a TypeScript orchestration script that routes the result through the Verdict Validator. Drivers with runnable public checkpoints also link the hermetic Modal job that loads the model and measures the paper claim.
- scripts/run-reproduction-qwen25.ts ↗— TS orchestrator (claim citation, DB write, validator wiring)
- modal/reproduce_qwen25.py ↗— Python Modal job (HF model load, evaluation harness, measurement)