System record
Human-tuned weak-to-strong method
Organization: Anthropic · base model: Qwen1.5-0.5B-Chat weak teacher, Qwen3-4B-Base strong student
System and setup
| Model version | not reported |
|---|---|
| Software used to run the model | Manual tuning by two authors |
| Software version | not reported |
| Tools available to the model | Four literature methods plus a zero-shot baseline considered |
| Release date | Not reported |
Autonomy assessments
No autonomy assessment is published for this system.
Published capability results
| Measure | Version | Reported result | Date |
|---|---|---|---|
| Automated weak-to-strong research study | April 2026 chat-preference study | 0.23 PGR | 2026-04 |
Related research
Additional system details
- System ID
- automated-w2s-human-tuned-method
- Organization
- Anthropic
- Base model
- Qwen1.5-0.5B-Chat weak teacher, Qwen3-4B-Base strong student
- Release date
- Not reported
- Assessment records
- 0
- Recursive evidence records
- 0