nl2bash_v1: validation run 2026-10-17-nl2bash_v1-b0c10dccb2cd

### Reproducibility and retries

| Stage | Mode | Tasks with expectation | Matched first attempt | Matched after 1 / 2 / 3+ retries | Archived as not robust | Not-run retried | Recovered | Unresolved |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | classifier | 0 | 0 | 0 / 0 / 0 | 0 | 0 | 0 | 0 |
| 2 | classifier | 0 | 0 | 0 / 0 / 0 | 0 | 0 | 0 | 0 |
| 3 | classifier | 0 | 0 | 0 / 0 / 0 | 0 | 0 | 0 | 0 |

Expected outcomes come from an earlier audit of the same tasks (matched by task ID). A task whose result differs from that audit is rerun up to the retry limit; a task that never matches is archived as not robust. Without an expectation only tasks with no result (crash, node failure, missing reward or missing runner evidence) are rerun, and a task still without a result after the last attempt is archived as not robust. Stages 1, 2 and 3 are covered; the retries run after the whole pipeline.

## nl2bash_v1

Keeps **1,425 tasks** and archives **72** from [open-thoughts/TaskTrove/nl2bash (TOML-normalized input)](https://huggingface.co/datasets/open-thoughts/TaskTrove) at commit `sha256:800161682dd6bc38194bad522e68f4c86edd0d247b992c8135de70b12f429d7b`. Patch version: `nl2bash-remove-unsupported-verifier-restart-v1`, `nl2bash-harness-seed-setup-v1`.

| Archive label | Tasks | Explanation |
| --- | ---: | --- |
| reference-solution-reward-zero | 72 | oracle reward [0.0] instead of 1 |

Our patcher modified **1,497 of the original 1,497 tasks**.

Automatic fixes changed **68 unique tasks**.

Tasks can have multiple change labels, so category counts may overlap.

| Change label | Tasks affected | Explanation |
| --- | ---: | --- |
| relative-path-fix | 68 | Replaced relative instruction paths with uniquely observed absolute paths. |
| repair:harness-seed-setup.py | 1497 | Renamed setup_files/setup_seeds.sh to setup_files/setup.sh and pointed solve.sh at it (it called the missing /tests/setup_seeds.sh), so the harness seeds /workspace before every trial; removed the instruction block that asked the agent to run the script; the fixtures line now says the input files are already in /workspace, and the claim that the verifier inspects helper files is removed (it reads only /output/command_capture.txt). |

## Runtime and validation time

| Setting | Recorded value |
| --- | --- |
| Cluster | helma/Slurm |
| Backend | apptainer |
| Architecture | x86_64 |
| Network | host |
| Python | 3.12.14 |
| Harbor | 0.8.1 |
| Requested CPUs | 32 |
| Trial concurrency | 17 |
| Static concurrency | 17 |
| terminal-bench commit | 1dcda8716784493721921c23e4bc7f7d988b4494 |
| harbor-validation commit | 7faf878c14b6d72737579ec936ebf5f74ba3194d |
| Unique cached images / tasks kept | 0 / 1,425 |
| Tasks with cached images | 0 / 1,425 |

| Data source | Unique upstream build environments | Unique kept build environments | Change |
| --- | ---: | ---: | ---: |
| nl2bash_v1 | 2 | 2 | +0 |


Build environments count distinct Dockerfile directory contents, file modes and links. Upstream covers the original source scope, including subsequently archived tasks; kept covers retained tasks. These are build contexts, not built image digests or cached-image counts.


Static checks excluded by the contract: 13. Recorded but not required: none. Reasons are in the run file.

Image manifest: `images/manifests/2026-10-17-nl2bash_v1-b0c10dccb2cd.json`. Cached images use Apptainer; Dockerfiles remain in the tasks.

Automatically generated data card:

- nl2bash_v1: `nl2bash_v1/README.md`

## Files and evidence

- Validation results and upstream comparison: `runs/2026-10-17-nl2bash_v1-b0c10dccb2cd.json`
- nl2bash_v1: dataset: `nl2bash_v1/tasks.parquet`
- nl2bash_v1: archived tasks: `nl2bash_v1/archive.parquet`
- nl2bash_v1: patch script: `evidence/2026-10-17-nl2bash_v1-b0c10dccb2cd/nl2bash_v1/patchers/0ff3d34f33f5d1d9678fdaf6b7567b85f8ee70126dbf2db7fc48206f9dad62de/patch.py`

evidence/2026-10-17-nl2bash_v1-b0c10dccb2cd/nl2bash_v1/patchers/0ff3d34f33f5d1d9678fdaf6b7567b85f8ee70126dbf2db7fc48206f9dad62de/patch.py

@@ -0,0 +1,177 @@
+"""Repair the NL2Bash verifier schema and let the harness run the seed setup."""
+
+import sys
+from pathlib import Path
+sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+
+import argparse
+import copy
+import gzip
+import hashlib
+import io
+import re
+import tarfile
+import tomllib
+
+REVISION = '946884702046be1a7dcea2638186ad6b0d2ea103'
+PATCH_VERSIONS = (
+    'nl2bash-remove-unsupported-verifier-restart-v1',
+    'nl2bash-harness-seed-setup-v1',
+)
+LABEL = 'static:task-toml-schema.py'
+SETUP_LABEL = 'repair:harness-seed-setup.py'
+REASONS = {
+    LABEL: ('Removed verifier.restart_environment from task.toml: the pinned Harbor '
+            'schema does not recognize this field and the runtime ignores it.'),
+    SETUP_LABEL: ('Renamed setup_files/setup_seeds.sh to setup_files/setup.sh and pointed '
+                  'solve.sh at it (it called the missing /tests/setup_seeds.sh), so the harness '
+                  'seeds /workspace before every trial; removed the instruction block that '
+                  'asked the agent to run the script; the fixtures line now says the input '
+                  'files are already in /workspace, and the claim that the verifier inspects '
+                  'helper files is removed (it reads only /output/command_capture.txt).'),
+}
+SEED_SCRIPT = 'setup_files/setup_seeds.sh'
+HARNESS_SCRIPT = 'setup_files/setup.sh'
+BROKEN_SETUP_CALL = b'bash /tests/setup_seeds.sh'
+HARNESS_SETUP_CALL = b'bash /setup_files/setup.sh'
+INSTRUCTION_SETUP = (b'## Environment Setup (run before starting)\n\n```bash\n'
+                     b'bash /setup_files/setup_seeds.sh\n```\n\n---\n\n')
+INPUT_LINES = (
+    b'- From `/workspace`, assemble any missing fixtures using standard shell utilities '
+    b'before running your command.\n',
+    b'- The workspace already contains pre-seeded fixtures copied to `/workspace`; '
+    b'run the command from that directory.\n',
+)
+INPUT_LINE = (b'- The input files for this task are already in `/workspace`; '
+              b'run the command from that directory.\n')
+HELPER_LINE = b'- Do not delete any helper files generated during execution; the verifier inspects them.\n'
+
+
+def rewrite_instruction(body):
+    """Drop the setup block and the two lines that misdescribe the seeded start state."""
+    text = body[len(INSTRUCTION_SETUP):]
+    if [text.count(line) for line in INPUT_LINES].count(1) != 1 or \
+            sum(text.count(line) for line in INPUT_LINES) != 1 or text.count(HELPER_LINE) != 1:
+        raise ValueError('instruction.md lacks the expected fixtures and helper-files lines')
+    for line in INPUT_LINES:
+        text = text.replace(line, INPUT_LINE)
+    return text.replace(HELPER_LINE, b'')
+
+
+def clean_config(content):
+    """Remove only the unsupported field; verify every other TOML value survives."""
+    text = content.decode('utf-8')
+    before = tomllib.loads(text)
+    if 'restart_environment' not in before.get('verifier', {}):
+        return content
+    expected = copy.deepcopy(before)
+    del expected['verifier']['restart_environment']
+    section = re.compile(r'(?ms)(^\[verifier\][^\n]*\n)(.*?)(?=^\[|\Z)')
+    field = re.compile(r'(?m)^[ \t]*(?:restart_environment|"restart_environment"|\'restart_environment\')[ \t]*=[^\n]*(?:\n|$)')
+    cleaned = section.sub(lambda m: m[1] + field.sub('', m[2]), text)
+    if tomllib.loads(cleaned) != expected:
+        raise ValueError('Unsupported task.toml layout for verifier.restart_environment removal')
+    return cleaned.encode('utf-8')
+
+
+def patch_archive(blob):
+    """Apply narrowly scoped fixes while preserving all other archive payloads."""
+    with tarfile.open(fileobj=io.BytesIO(blob), mode='r:*') as source:
+        members = source.getmembers()
+        configs = [m for m in members if Path(m.name).as_posix() == 'task.toml']
+        if len(configs) != 1 or not configs[0].isfile():
+            raise ValueError('Expected exactly one regular task.toml in task archive')
+        config = configs[0]
+        original = source.extractfile(config).read()
+        cleaned = clean_config(original)
+        edited = {}
+        labels = []
+        if cleaned != original:
+            edited[id(config)] = cleaned
+            labels.append(LABEL)
+
+        by_name = {Path(m.name).as_posix(): m for m in members}
+        helper = by_name.get(SEED_SCRIPT)
+        solutions = [m for m in members
+                     if Path(m.name).as_posix().startswith('solution/') and m.isfile()]
+        solution_bodies = {id(m): source.extractfile(m).read() for m in solutions}
+        instruction = by_name.get('instruction.md')
+        instruction_body = source.extractfile(instruction).read() if instruction else b''
+        calls = any(BROKEN_SETUP_CALL in body for body in solution_bodies.values())
+        renamed = {}
+        if helper is not None or calls or b'setup_seeds' in instruction_body:
+            # All three parts belong together; anything else needs a human look.
+            if helper is None or not helper.isfile() or not calls or HARNESS_SCRIPT in by_name:
+                raise ValueError('Unexpected seed setup layout: need setup_files/setup_seeds.sh, '
+                                 'a /tests/setup_seeds.sh call in solve.sh and no setup_files/setup.sh')
+            if not instruction_body.startswith(INSTRUCTION_SETUP) or \
+                    b'setup_seeds' in instruction_body[len(INSTRUCTION_SETUP):]:
+                raise ValueError('instruction.md does not start with the expected seed setup block')
+            renamed[id(helper)] = HARNESS_SCRIPT
+            edited[id(instruction)] = rewrite_instruction(instruction_body)
+            for member in solutions:
+                body = solution_bodies[id(member)]
+                fixed = body.replace(BROKEN_SETUP_CALL, HARNESS_SETUP_CALL)
+                if fixed != body:
+                    edited[id(member)] = fixed
+            labels.append(SETUP_LABEL)
+        if not edited:
+            return blob, []
+        output = io.BytesIO()
+        compressed = blob[:2] == b'\x1f\x8b'
+        stream = gzip.GzipFile(filename='', fileobj=output, mode='wb', mtime=0) if compressed else output
+        with tarfile.open(fileobj=stream, mode='w', format=source.format) as target:
+            for member in members:
+                info = copy.copy(member)
+                info.name = renamed.get(id(member), member.name)
+                body = edited.get(id(member), source.extractfile(member).read()) if member.isfile() else None
+                if member.isfile() and id(member) in edited:
+                    info.size = len(body)
+                target.addfile(info, io.BytesIO(body) if body is not None else None)
+        if compressed:
+            stream.close()
+        return output.getvalue(), labels
+
+
+def patch(source, output):
+    import pyarrow as pa
+    import pyarrow.parquet as pq
+    from data.utils.patch_reporting import write_patch_report
+
+    source, output = Path(source), Path(output)
+    output.mkdir(parents=True, exist_ok=True)
+    destination = output / 'tasks.parquet'
+    if destination.exists():
+        raise ValueError('Use a new patch output directory')
+    parquet = pq.ParquetFile(source)
+    changed = {}
+    with pq.ParquetWriter(destination, parquet.schema_arrow) as writer:
+        for batch in parquet.iter_batches(batch_size=32):
+            rows = batch.to_pylist()
+            for row in rows:
+                row['task_binary'], labels = patch_archive(row['task_binary'])
+                if labels:
+                    changed[row['path']] = labels
+            writer.write_table(pa.Table.from_pylist(rows, schema=parquet.schema_arrow))
+    with source.open('rb') as source_stream:
+        source_hash = hashlib.file_digest(source_stream, 'sha256').hexdigest()
+    return write_patch_report(source, destination, source={
+        'dataset': 'open-thoughts/TaskTrove/nl2bash (TOML-normalized input)',
+        'url': 'https://huggingface.co/datasets/open-thoughts/TaskTrove',
+        'revision': f'sha256:{source_hash}'}, dropped={}, change_labels=changed,
+        change_reasons={name: '; '.join(REASONS[label] for label in labels)
+                        for name, labels in changed.items()},
+        patches=PATCH_VERSIONS, patcher=__file__)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--source', type=Path, required=True)
+    parser.add_argument('--output', type=Path, required=True)
+    args = parser.parse_args()
+    report = patch(args.source, args.output)
+    print(f"NL2Bash: {sum(t['action'] == 'changed' for t in report['tasks'])} tasks repaired")
+
+
+if __name__ == '__main__':
+    main()

nl2bash_v1/README.md

@@ -0,0 +1,34 @@
+# nl2bash_v1
+
+## Original datasource
+
+- NL2Bash — https://arxiv.org/abs/1802.08979
+
+## Domains
+
+File Operations, System Administration
+
+- **File Operations:** Every sampled task asks the agent to locate, filter, count, sort, or copy files in a seeded /workspace using find, grep, xargs, du, sort, cp and friends (e.g. find .log files up to depth 3 and list the three largest; copy file.ext matches with attributes preserved; count zero-byte files under subdir), then write the command's combined stdout/stderr to /output/command_capture.txt.
+- **System Administration:** Tasks are framed as 'Terminal Automation Requests' in an Ubuntu container and are solved by composing shell one-liners with coreutils, bash variables, redirection and pipelines (e.g. capturing whoami into a variable, 2>&1 capture, -print0/xargs -0 handling); the manifests show each task was generated from an NL2Bash NL-to-bash pair with added flags such as -maxdepth, -type f, -iname.
+
+## Agent capabilities
+
+Specification Adherence
+
+- **Specification Adherence:** Intended practice: each goal stacks several precise, easily-dropped details (depth limit, case-insensitivity, regular-files-only, exact extension, numeric sort on a specific field, zero matches omitted, capture of both stdout and stderr to a fixed path). The verifier compares a normalized multiset of output lines against the reference command's output and fails on any missing record or error text, so dropping one qualifier changes the output and zeroes the reward. No trajectories were supplied, so this is the designed target, not observed behavior.
+
+## Expected benchmark transfer
+
+These are hypotheses, not measured training gains.
+
+- **Terminal-Bench 2.0**: The tasks use the same Harbor-style layout (task.toml, tests/test.sh writing /logs/verifier/reward.txt, /output artifacts) and train the core skill of translating a natural-language request into correct shell pipelines inside an Ubuntu container, which is a prerequisite for most TB2 tasks even though TB2 tasks are far longer-horizon.
+- **OpenThoughts-TBLite 2.0**: TBLite is a faster, lighter terminal-agent panel calibrated to Terminal-Bench 2; practice at precise find/grep/sort/xargs command composition and writing results to a required output path should transfer directly to its shorter shell-centric tasks.
+- **Harbor-Index**: Harbor-Index is a Harbor-format terminal-agent panel; the dataset's tasks are built in the same task.toml/verifier format and exercise basic Linux file-manipulation commands with executable verification, so modest gains on its shell-heavy tasks are plausible.
+
+## Files
+
+- `tasks.parquet`: kept tasks.
+- `archive.parquet`: excluded tasks with validation reasons.
+
+Generated with claude-fable-5-1; domain and capability labels follow CLI-Universe.
+Taxonomy source: https://arxiv.org/abs/2606.22883v1

nl2bash_v1/archive.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ca84631bfe4f8c2261449b005548478fe92997edbb0019fe724c4b90c1bcabd
+size 330636

nl2bash_v1/tasks.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c374c3be62789938fa4930323ff4054c45b062c254898e875ca61aaf8ba980e
+size 4859259