# NL2Bash v2

## Dataset stats

- **Original tasks:** 1,497
- **Kept:** 938
- **Changed from v1:** 0 task payloads
- **Archived:** 559 (487 with no Coder success and 72 earlier v1 exclusions)
- **Pending attempts:** 0
- **Evaluated tasks:** 1,425
- **Recorded attempts:** 5,700
- **Saved conversations:** 5,700
- **Successful attempts:** 2,911
- **Images:** 1 before, 1 now

Based on [nl2bash_v1](https://huggingface.co/datasets/FWeindel/validated-tasks/tree/3abfc63260ef8ab59ac61274450310f650585bc9/nl2bash_v1) at revision `3abfc63260ef8ab59ac61274450310f650585bc9`.

## Evaluation settings

Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget.

| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling |
| --- | ---: | ---: | --- | --- | --- |
| openai/coder-30b | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} |

Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults.

**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited.

The table reports settings found in **1 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results.

## Dataset characteristics and changes from v1

[NL2Bash](https://arxiv.org/abs/1802.08979) tasks from [TaskTrove](https://huggingface.co/datasets/open-thoughts/TaskTrove) `nl2bash` at revision `946884702046be1a7dcea2638186ad6b0d2ea103`.

<details><summary>Domains</summary>

**File Operations:** locate, filter, count, sort or copy files in a seeded `/workspace` with find, grep, xargs, du, sort and cp, then capture the command's output in `/output/command_capture.txt`. **System Administration:** compose shell one-liners with coreutils, variables, redirection and pipelines in an Ubuntu container.

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

**Specification Adherence:** each request stacks precise details (depth limit, case-insensitivity, regular files only, numeric sort, capturing both stdout and stderr), and the verifier compares the output lines with the reference command's output, so dropping one detail fails the task. Labels follow [CLI-Universe](https://arxiv.org/abs/2606.22883v1).

</details>

Intended benchmark transfer: **Terminal-Bench 2.0, OpenThoughts-TBLite 2.0 and Harbor-Index**, through translating requests into correct shell pipelines in a Harbor-style container.

Kept 938 tasks solved at least once in four attempts. Task payloads and the pinned v1 image are unchanged. Archived 487 tasks with zero observed successes under this evaluation, preserving their payloads and attempt evidence. Earlier archive reasons are preserved for all 72 v1 exclusions.

The v1 release documents build/setup, reference-solution, no-op and rule-based validation. Its Git-history and timestamp leak checks were not run. This release filters by teacher results and does not establish completion of those outstanding validation checks.

## Evaluation setup and coverage

| Setting | Value |
| --- | --- |
| Teacher | Qwen/Qwen3-Coder-30B-A3B-Instruct |
| Agent | Terminus 2 |
| Coverage | 1,425 tasks, four attempts each, 5,700 total |
| Sampling | Temperature 0.7; top-p 0.8; top-k 20; repetition penalty 1.05 |
| Token budgets | 32,768 context; intended 8,192 output tokens per call (see enforcement caveat) |
| Serving | One H200 per shard; eight concurrent trials; 32 CPUs, including four serving CPUs |
| Evidence | 61 complete stage-6 reports and 61 complete stage-7 reports, including the disjoint eight-task pilot |

The 938 kept tasks each declare a 600-second agent timeout and a 600-second verifier timeout. [Frozen contracts and provenance](metadata/provenance.json) record model revision, execution settings, task hashes and input evidence. Only the replacement pilot and full-array results are included.

## Pass rates and success distribution

All rates cover the full 1,425-task evaluated set, before filtering. Higher-k values use the saved four-attempt samples.

| Metric | Rate |
| --- | ---: |
| pass@1 | 51.07% |
| pass@2 | 59.79% |
| pass@3 | 63.65% |
| pass@4 | 65.82% |

| Successful attempts out of four | Tasks |
| ---: | ---: |
| 0 | 487 |
| 1 | 124 |
| 2 | 144 |
| 3 | 181 |
| 4 | 489 |

Observed reward groups: 489 always solved, 487 always zero, and 449 varying. These describe the recorded samples, not intrinsic task difficulty.

![Pass rates](plots/pass-rates.png)

![Successful attempts per task](plots/solved-counts.png)

## Trajectories and trace endings

All 5,700 saved conversations are included in [traces/coder.parquet](traces/coder.parquet), with original result records, task and attempt IDs, raw verifier rewards, effective rewards and the `retained_v2` flag. No conversations were filtered out of this evidence export. Of these, 2,911 attempts received verifier reward 1.

There are 5,699 verifier grades and one ungraded terminal `TmuxSessionEndedError`. Following the recorded `terminal-errors-zero-v2` policy, its effective reward is zero; its raw `verifier_reward` stays null. Its task has no success in the other three attempts and is archived. No absent attempts were filled in.

Trace endings: 4,249 task-complete and 1,451 tmux-session-closed. Session closure and grading are separate: 1,450 of those closures still have a verifier grade. Stage 7 completed for all tasks; presence of a conversation is not an additional training-quality acceptance claim.

![Trace outcomes](trace_stats/plots/all-models-trace-outcomes.png)

## Runtime, tokens and resources

| Measurement | Attempts measured | Median | Maximum |
| --- | ---: | ---: | ---: |
| Agent turns | 5,700 | 11 | 497 |
| Input tokens per attempt | 5,700 | 25,332 | 6,915,919 |
| Output tokens per attempt | 5,700 | 1,529 | 54,555 |
| Largest recorded turn, input plus output tokens | 5,700 | 3,625 | 35,245 |
| Agent time, seconds | 5,700 | 17.31 | 554.40 |
| Verifier time, seconds | 5,699 | 5.87 | 12.47 |
| Trial time, seconds | 5,700 | 31.71 | 568.58 |

Input and output totals sum model calls within each attempt. Recorded input-plus-output turn size is a trace measurement, not a configured serving context limit. Per-task peak memory is not reported here.

![Trace distributions](trace_stats/plots/all-models-trace-distributions.png)

See [trajectory statistics](trace_stats/README.md) for distributions, command counts, error categories and machine-readable per-attempt metrics.

## Files and selection

- [tasks.parquet](tasks.parquet): the 938 tasks solved at least once, with `coder_successes`, `coder_attempts` and `coder_graded_attempts` columns.
- [archive.parquet](archive.parquet): 559 excluded tasks, preserving earlier exclusion reasons. The 72 earlier exclusions have null Coder counts because they were not evaluated.
- [selection evidence](metadata/selection.json): per-task success counts and raw attempt grades.
- [trace manifest](traces/manifest.json): trace coverage, reward policy and file checksum.
- [patch.py](patch.py): reproduces the kept task file from the pinned v1 source and saved selection.

The task Parquet preserves pinned references to the existing v1 image bundle; no image rebuild is needed for this selection.