| name | compare-models |
| description | Compare results across models (Qwen vs Llama) at matching compression configurations. Generates side-by-side tables and identifies cross-model patterns. |
| argument-hint | ["num-prompts"] |
| allowed-tools | Read, Grep, Glob, Bash |
Compare Models
Generate a comprehensive cross-model comparison for all matching compression configurations.
Steps
-
Find matching result files for both models:
find results -name "*Qwen*_${0:-500}p.json" | sort
find results -name "*Llama*_${0:-500}p.json" | sort
-
For each press × ratio, load both models' results and extract:
- Accuracy (% correct GSM8K answers)
- CFR (looping + non-termination rate)
- Individual catastrophe rates (looping, non-termination, wrong_answer)
- Mean/max entropy, min top-1 probability
- Average tokens generated
-
Generate comparison tables:
Table: Accuracy Comparison
| Press | Ratio | Qwen 7B | Llama 8B | Delta |
|---|
Table: CFR Comparison
| Press | Ratio | Qwen CFR | Llama CFR | Delta |
|---|
Table: Failure Mode Comparison
| Press | Ratio | Qwen Loop% | Llama Loop% | Qwen NT% | Llama NT% |
|---|
-
Identify patterns:
- Which model is more robust to compression overall?
- Do they fail at the same compression ratios or different ones?
- Are failure modes the same (both loop) or different (one loops, one non-terminates)?
- Where does the "cliff" occur for each model × compressor?
-
Cross-model transfer implications:
- If failure signatures are similar across models → predictor may transfer
- If failure modes differ → per-model predictors may be needed
- This directly informs Phase 3 (P3-04: Cross-model transfer experiment)