# TermiGen v2

## Dataset stats

- **Tasks:** 3,566 original; 3,336 in v1; **2,277 kept in v2**, all previously patched. 1,059 v1 tasks have no recorded successful attempt; 230 earlier exclusions remain in v1.
- **Images:** reuse v1 environments and its cached images as they become available. No additional images or task repairs in v2.

## Dataset characteristics

[TermiGen](https://github.com/ucsb-mlsec/terminal-bench-env/tree/03bddac74ada2fa344df0e93b4a3ace2034294d1) supplies terminal tasks with prepared files and executable verifiers.

<details><summary>Domains</summary>

Software Engineering, Debugging, System Administration, Security, Data Processing, Scientific Computing and Machine Learning: repair code, diagnose services, inspect artifacts and configure computational workflows.

</details>
<details><summary>Capabilities</summary>

Exploration, Error Recovery and Specification Adherence; specialized tasks also involve Reverse Engineering and Adversarial Reasoning. Labels follow [CLI-Universe](https://arxiv.org/abs/2606.22883v1).

</details>

**Intended benchmark transfer:** terminal problem solving for Terminal-Bench; debugging and environment diagnosis for repository-based software engineering.

## Selection

Keep tasks with at least one recorded reward of 1 from either model, with four attempts per task and model.

| Model | Tasks solved at least once |
| --- | ---: |
| Qwen3-Coder-30B-A3B-Instruct | 2,123 / 3,336 |
| Qwen3-30B-A3B-Instruct-2507 | 1,850 / 3,336 |
| Either model | 2,277 / 3,336 |

All 3,336 evaluated task hashes match v1. V2 preserves task payloads exactly. Build, no-op and rule-based checks carry over from [v1](../termigen_v1/README.md); no reference solutions are provided.

Missing rewards remain unknown, not failures (142 Coder attempts and 179 Instruct attempts). Excluded tasks are those without a recorded success, not confirmed defective tasks.

### Successful attempts among kept tasks

| Instruct successes out of 4 | Tasks |
| --- | ---: |
| 0 | 427 |
| 1 | 417 |
| 2 | 359 |
| 3 | 421 |
| 4 | 653 |

Per-task Coder evidence establishes whether a task was solved; exact per-task success counts are not available in the published report, so they are not inferred.

## Files and selecting tasks

`tasks.parquet` contains kept tasks, `coder_solved`, `instruct_successes`, `instruct_missing_rewards` and `attempts_per_model`. `metadata/selection.json` records every v1 task's selection and pins the source revision. Excluded payloads remain in that v1 revision; no duplicate archive Parquet is uploaded. `patch.py` reproduces this selection.

```python
import pyarrow.parquet as pq
subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)])
pq.write_table(subset, "selected.parquet", compression="zstd")
```

Saved teacher trajectories from [PR #20](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/20) are available in [traces/](traces/), with coverage and provenance in [traces/manifest.json](traces/manifest.json). The 70 Parquet shards contain 26,688 attempt records from both models across all 3,336 evaluated v1 tasks. Filter `retained_v2 == True` for the 2,277 tasks kept in v2. The manifest records 64 missing trajectories and 321 ungraded attempts; missing raw verifier rewards remain null. It also documents missing-reward metadata discrepancies for 103 tasks.

## Runtime

Apptainer 1.5.4, Python 3.12.7 / 3.12.14, Harbor 0.8.1; [Marin Harbor bridge](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) and [runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). Teacher runs used Terminus 2; [recorded model results](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/11).