# EndlessTerminals v2

## Dataset statistics

- **2,492 original tasks → 2,144 repaired v1 tasks → 1,323 kept in v2.** All kept tasks were patched in v1; v2 changes no task payloads.
- **821 v1 tasks excluded for now:** 818 had no successful attempt; three have no recorded results. Excluded payloads remain in [v1](../endless_v1).
- **2,492 original build contexts → 1,834 cached Apptainer environments in v1.** V2 reuses **1,159** of those environments; no duplicate image upload.

## Dataset characteristics

<details><summary>Domains</summary>

**System Administration:** shell scripts, services and configuration. **File Operations:** exact placement, links, backups and compression. **Data Processing:** parse CSV, INI and network summaries into precise reports.

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

**Exploration, Specification Adherence and Constraint Satisfaction:** inspect supplied files and produce exact content, paths, permissions and formats. Labels follow [CLI-Universe](https://arxiv.org/abs/2606.22883v1).

</details>

Intended benchmark transfer: **Terminal-Bench 2.0 and OpenThoughts-TBLite 2.0**, through terminal inspection, shell scripting and filesystem tasks.

## Validation and filtering

- **Environment reliability and rule-based task quality:** keeps all [v1 repairs](../endless_v1), including build/setup, reference-solution and no-op verification, and static-check repairs.
- **Pass@4 filtering:** retain tasks solved at least once by **Qwen3-30B-A3B-Instruct-2507**, using Terminus 2. Evaluated content hashes match every kept task.
- **15 October 2026 results:** 2,141 tasks completed four attempts each; none are partial; three have no recorded attempts. Terminal execution failures count as unsuccessful attempts. Five attempts have no raw verifier reward and receive a filtering score of zero.

### Successful attempts among kept tasks

| Successful attempts | Tasks |
| --- | ---: |
| 1/4 | 288 |
| 2/4 | 254 |
| 3/4 | 261 |
| 4/4 | 520 |

### Select by successful attempts

```python
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

path = hf_hub_download("FWeindel/validated-tasks", "endless_v2/tasks.parquet", repo_type="dataset")
tasks = pq.read_table(path, filters=[("instruct_successes", ">=", 1), ("instruct_successes", "<=", 2)])
```

`metadata/selection.json` records all v1 task decisions and counts. `patch.py` reproduces the task Parquet from pinned v1 data. `instruct_graded_attempts` counts attempts assigned a filtering score, including terminal failures.

## Traces

[Compressed traces](traces/attempts.parquet) include **all 8,564 attempts**, successful and failed, across kept and excluded tasks. No conversations are missing. Each row contains the task ID, model, attempt ID, filtering reward, raw verifier reward, retention flag and full trajectory and result JSON.

## Resources

Historical teacher runs used **one H200, 32 CPUs per job and concurrency 10**. Fifteen jobs each took approximately **55 to 79 minutes**. Environment and image versions are inherited from [v1](../endless_v1).