metadata
license: mit
tags:
- strategy_compliance
- musr
- murder_mysteries
- reasoning_malleability
- baseline
t1-strategy-musr-baseline-qwen3-80b-instruct-musr-base
Strategy compliance evaluation on MuSR murder mysteries — baseline variant.
Model received standard MuSR cot+ prompt (baseline control). Judge still scores against criterion-first rubric. Compliance is scored by an LLM judge (1-5 Likert) against the Criterion-First rubric.
Results
| Metric | Value |
|---|---|
| pass@1 | 0.8000 |
| Strategy compliance (mean) | 2.90 |
| Strategy compliance (min) | 2 |
| Strategy compliance (max) | 4 |
| Total problems | 10 |
Compliance Distribution
| Score | Count | Meaning |
|---|---|---|
| 1 | 0 | Non-compliant |
| 2 | 5 | Minimal |
| 3 | 1 | Partial |
| 4 | 4 | Mostly compliant |
| 5 | 0 | Fully compliant |
Details
| Parameter | Value |
|---|---|
| Model | together_ai/Qwen/Qwen3-Next-80B-A3B-Instruct |
| Thinking | False |
| Temperature | 0.7 |
| Max tokens | 32768 |
| Judge model | openai/gpt-4o-mini |
| Num examples | 10 |
| Variant | baseline |
Usage
from datasets import load_dataset
ds = load_dataset("reasoning-degeneration-dev/t1-strategy-musr-baseline-qwen3-80b-instruct-musr-base", split="train")
scores = ds["strategy_compliance"]
print(f"Mean compliance: {sum(scores) / len(scores):.2f}")
Tracked in reasoning-degeneration-dev/PROJECT-MANIFEST