Terminal-Bench 2.1terminalPython
Recover and tune a PyTorch model architecture
Reconstruct the network represented by a state dictionary, freeze all but output_layer, improve dataset MSE, and export the result as TorchScript.
pytorch-model-recovery
Normalized profile
What this task asks for
Reconstruct the network represented by a state dictionary, freeze all but output_layer, improve dataset MSE, and export the result as TorchScript. Define RecoveredModel to match /app/weights.pt exactly. Leave /app/weights.pt unchanged. Modify only output_layer parameters. Reduce MSE on /app/dataset.pt. Save a TorchScript model compatible with the original state dictionary.
Intent
- feature implementation
- model training
- test and validation
Work surfaces
- model architecture
- state dictionary
- training loop
Technology
- PyTorch
- TorchScript
- transformers
Expected artifacts
- /app/model.pt
What makes it difficult
- reverse engineering from incomplete or indirect evidence
- strict behavioral compatibility or exact-output validation
- changes span multiple components or interfaces
Published evidence
Model × harness outcomes
Counts are tied to these exact configurations. Harnesses are not treated as equivalent across benchmarks.
- Models
- 13
- Configurations
- 20
- Trials
- 100
Showing 20 of 20 configurations
| Model | Harness | Effort | Result | Passes |
|---|---|---|---|---|
| claude-fable-5anthropic · 2026-06-07 | claude-code2.1.167 | xhigh | Mixed1 errors | 4/5 |
| claude-fable-5anthropic · 2026-06-05 | terminus-22.0.0 | high | Mixed1 errors | 4/5 |
| claude-opus-4-7anthropic · 2026-05-01 | claude-code2.1.123 | max | Mixed3 errors | 2/5 |
| claude-opus-4-7anthropic · 2026-05-01 | terminus-22.0.0 | max | Mixed1 errors | 4/5 |
| claude-opus-4-8anthropic · 2026-07-09 | claude-code2.1.205 | high | All passed | 5/5 |
| claude-sonnet-5anthropic · 2026-07-09 | claude-code2.1.205 | high | All passed | 5/5 |
| gemini-3-pro-previewgemini · 2026-05-01 | gemini-cli0.40.0 | high | All passed | 5/5 |
| gemini-3-pro-previewgemini · 2026-05-01 | terminus-22.0.0 | high | All passed | 5/5 |
| gemini-3.1-pro-previewgemini · 2026-05-05 | gemini-cli0.40.0 | high | Mixed3 errors | 2/5 |
| gemini-3.1-pro-previewgemini · 2026-05-05 | terminus-22.0.0 | high | All passed | 5/5 |
| glm-5.1zai · 2026-05-01 | claude-code2.1.123 | max | Mixed3 errors | 2/5 |
| gpt-5.5openai · 2026-05-01 | codex0.125.0 | xhigh | Mixed | 3/5 |
| gpt-5.5openai · 2026-05-01 | terminus-22.0.0 | xhigh | All passed | 5/5 |
| gpt-5.6-lunaopenai · 2026-07-10 | codex0.144.0 | max | All passed | 5/5 |
| gpt-5.6-lunaopenai · 2026-07-11 | codex0.144.1 | max | All passed | 5/5 |
| gpt-5.6-solopenai · 2026-07-10 | codex0.144.0 | max | Mixed | 1/5 |
| gpt-5.6-terraopenai · 2026-07-10 | codex0.144.0 | max | All passed | 5/5 |
| gpt-5.6-terraopenai · 2026-07-11 | codex0.144.1 | max | All passed | 5/5 |
| grok-4.5cursor · 2026-07-09 | cursor-cli2026.07.08-0c04a8a | default | All passed | 5/5 |
| muse-spark-1.1openai · 2026-07-09 | mini-swe-agent2.4.5 | xhigh | All passed | 5/5 |