nl2bash_v1: reviewed dataset card; fix task_267 output path in the goal

Card: patches, per-stage results, instruction-path review and causes of the 72 archived tasks. task_267: the goal named $PWD/command_capture.txt (resolves to /workspace) while the verifier reads /output/command_capture.txt; one-sentence instruction change after validation, content_sha256 recomputed.

nl2bash_v1/README.md

@@ -1,4 +1,9 @@
-# nl2bash_v1
+# NL2Bash
+
+## Dataset stats
+
+- **Tasks:** 1,497 original, 1,425 kept, all 1,497 patched; 72 archived.
+- **Source:** [TaskTrove](https://huggingface.co/datasets/open-thoughts/TaskTrove) `nl2bash` at revision `946884702046be1a7dcea2638186ad6b0d2ea103`.
 
 ## Original datasource
 
@@ -9,7 +14,7 @@
 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.
+- **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
 
@@ -25,10 +30,38 @@ These are hypotheses, not measured training gains.
 - **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.
 
+## Patches
+
+- **Verifier schema:** removed `verifier.restart_environment` from `task.toml`; the pinned Harbor schema does not recognize it and the runtime ignores it (1,497 tasks).
+- **Seed setup:** renamed `setup_files/setup_seeds.sh` to `setup_files/setup.sh`, so the harness runs it automatically before the agent and the verifier start (the reference solution called a script that did not exist). Removed the instruction block that told the agent to run it, and the instructions now say the input files are already in `/workspace`. Also removed the claim that the verifier inspects helper files: it reads only `/output/command_capture.txt` (1,497 tasks).
+- **Output path (task_267):** the goal said to capture output in `$PWD/command_capture.txt`, where `$PWD` resolves to `/workspace`, but the verifier expects the output in `/output/command_capture.txt`. The goal now names `/output/command_capture.txt`.
+
+## Per-stage breakdown
+
+- **Build and setup:** 1,497 passed, 0 archived.
+- **Reference solution (reward 1):** 1,425 passed, 72 archived (see below).
+- **No-op baseline (reward 0):** 1,497 passed, 0 archived.
+- **Rule-based checks:** 1,497 passed; 111 advisory path warnings (see below).
+
+## Notes
+
+**Instruction paths.** The automatic path check flagged relative paths in 111 tasks. We inspected all of them: 110 are clear to the agent. In 96 the path is a seed folder that setup copies into `/workspace`, the directory the instructions say to work from; in 24 the task itself creates the path; one is an exclusion pattern, and three were filename patterns rather than paths. One task was misleading and is fixed (task_267, see Patches).
+
+**Archived: reference solution scores 0 (72 tasks).** The verifier compares the output with an `expected_output` recorded in the original environment, which cannot be reproduced in this environment:
+
+- **Fixtures changed in the TaskTrove conversion:** e.g. a 50,000-line input became a zero-filled file (task_2706); files meant to be older than 10 days are created fresh, so `find -atime +10` finds nothing (task_9135).
+- **Output tied to the recording machine:** kernel version, hostname, date or file timestamps.
+- **Output that depends on the filesystem:** `find` order, `du` sizes, filesystem type.
+- **Output that depends on timing:** commands that read the capture file they are writing.
+
+## Runtime
+
+- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**.
+- **Resources:** Helma; 48 CPU cores, 64 GiB RAM per job, 17 concurrent tasks, no GPUs.
+
 ## 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
+Domain and capability labels follow CLI-Universe. Taxonomy source: https://arxiv.org/abs/2606.22883v1

nl2bash_v1/tasks.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9c374c3be62789938fa4930323ff4054c45b062c254898e875ca61aaf8ba980e
-size 4859259
+oid sha256:ea08f8cf7f49d8472bc515a3c0a674b8c21f199a24f12c086301b6480958d052
+size 855485