Improve dataset card and project links
Browse files
README.md
CHANGED
|
@@ -5,22 +5,43 @@ tags:
|
|
| 5 |
- paper-reproduction
|
| 6 |
- execution-trajectories
|
| 7 |
- reproducibility
|
|
|
|
|
|
|
| 8 |
size_categories:
|
| 9 |
- n<1K
|
| 10 |
---
|
| 11 |
|
| 12 |
# ReproBoundary Execution Trajectories
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
This dataset contains the released execution logs for ReproBoundary, a
|
| 15 |
benchmark of agent capability boundaries in complex paper-to-code
|
| 16 |
reproduction. The benchmark defines 47 fixed paper-reproduction tasks and runs
|
| 17 |
each task with nine model-agent settings, producing 423 trajectories.
|
| 18 |
|
| 19 |
-
The task definitions, evaluator code, protocol templates, and website
|
| 20 |
-
in the [ReproBoundary GitHub repository](https://github.com/
|
| 21 |
This dataset repository stores the larger trajectory files and the small
|
| 22 |
metadata needed to interpret and verify them.
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
## Contents
|
| 25 |
|
| 26 |
- `trajectories/TRAJECTORY_INDEX.csv` links every run to its task,
|
|
@@ -34,6 +55,17 @@ metadata needed to interpret and verify them.
|
|
| 34 |
- `DATASET_SHA256SUMS` fixes the bytes of every released file.
|
| 35 |
- `scripts/verify_dataset.py` verifies the dataset after download.
|
| 36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
## Download
|
| 38 |
|
| 39 |
This dataset is public. Install the official Hugging Face client in an isolated
|
|
@@ -50,9 +82,10 @@ hf download SupingSun/ReproBoundary-Execution-Trajectories \
|
|
| 50 |
--local-dir ReproBoundary-Execution-Trajectories
|
| 51 |
|
| 52 |
cd ReproBoundary-Execution-Trajectories
|
|
|
|
| 53 |
```
|
| 54 |
|
| 55 |
-
The
|
| 56 |
of free space for the dataset, client environment, and temporary files.
|
| 57 |
|
| 58 |
## Recorded run states
|
|
|
|
| 5 |
- paper-reproduction
|
| 6 |
- execution-trajectories
|
| 7 |
- reproducibility
|
| 8 |
+
- benchmark
|
| 9 |
+
- research-agents
|
| 10 |
size_categories:
|
| 11 |
- n<1K
|
| 12 |
---
|
| 13 |
|
| 14 |
# ReproBoundary Execution Trajectories
|
| 15 |
|
| 16 |
+
[Benchmark website](https://reproboundary.github.io/) ·
|
| 17 |
+
[GitHub release](https://github.com/ReproBoundary/ReproBoundary) ·
|
| 18 |
+
[Task inventory](https://github.com/ReproBoundary/ReproBoundary/blob/main/task_pool/TASK_INVENTORY_47.csv) ·
|
| 19 |
+
[Evaluator guide](https://github.com/ReproBoundary/ReproBoundary/blob/main/docs/EVALUATOR_PROVENANCE.md)
|
| 20 |
+
|
| 21 |
This dataset contains the released execution logs for ReproBoundary, a
|
| 22 |
benchmark of agent capability boundaries in complex paper-to-code
|
| 23 |
reproduction. The benchmark defines 47 fixed paper-reproduction tasks and runs
|
| 24 |
each task with nine model-agent settings, producing 423 trajectories.
|
| 25 |
|
| 26 |
+
The task definitions, evaluator code, protocol templates, and website source are
|
| 27 |
+
kept in the [ReproBoundary GitHub repository](https://github.com/ReproBoundary/ReproBoundary).
|
| 28 |
This dataset repository stores the larger trajectory files and the small
|
| 29 |
metadata needed to interpret and verify them.
|
| 30 |
|
| 31 |
+
> **What this is:** a provenance-indexed archive of agent execution logs.
|
| 32 |
+
> **What this is not:** a package of checkpoints, predictions, model weights,
|
| 33 |
+
> training states, or independently verified reproduction outcomes.
|
| 34 |
+
|
| 35 |
+
## At a glance
|
| 36 |
+
|
| 37 |
+
| Released material | Count | Role |
|
| 38 |
+
| --- | ---: | --- |
|
| 39 |
+
| Paper-reproduction tasks | 47 | One frozen target experiment per source paper |
|
| 40 |
+
| Model-agent settings | 9 | The evaluated model and coding-agent combinations |
|
| 41 |
+
| Execution trajectories | 423 | One cleaned `agent.log.gz` file per planned run |
|
| 42 |
+
| Completed evidence bundles | 418 | Recorded completion state, not scientific success |
|
| 43 |
+
| Unresolved or active snapshots | 5 | Preserved without guessing a final state |
|
| 44 |
+
|
| 45 |
## Contents
|
| 46 |
|
| 47 |
- `trajectories/TRAJECTORY_INDEX.csv` links every run to its task,
|
|
|
|
| 55 |
- `DATASET_SHA256SUMS` fixes the bytes of every released file.
|
| 56 |
- `scripts/verify_dataset.py` verifies the dataset after download.
|
| 57 |
|
| 58 |
+
Hugging Face's row viewer does not render the compressed logs as a tabular
|
| 59 |
+
dataset. Start with `trajectories/TRAJECTORY_INDEX.csv` to browse run metadata,
|
| 60 |
+
then open only the corresponding gzip files needed for an analysis.
|
| 61 |
+
|
| 62 |
+
### Index fields
|
| 63 |
+
|
| 64 |
+
The central trajectory index records the task, model-agent setting, preserved
|
| 65 |
+
run state, released log path, byte size, redaction count, and SHA-256 file
|
| 66 |
+
fingerprint. Task definitions and paper-reported reference values live in
|
| 67 |
+
`task_pool/tasks.jsonl` and `task_pool/TASK_INVENTORY_47.csv`.
|
| 68 |
+
|
| 69 |
## Download
|
| 70 |
|
| 71 |
This dataset is public. Install the official Hugging Face client in an isolated
|
|
|
|
| 82 |
--local-dir ReproBoundary-Execution-Trajectories
|
| 83 |
|
| 84 |
cd ReproBoundary-Execution-Trajectories
|
| 85 |
+
python3 scripts/verify_dataset.py
|
| 86 |
```
|
| 87 |
|
| 88 |
+
The fixed release snapshot is about 399 MB and contains 423 gzip logs. Allow at least 1 GB
|
| 89 |
of free space for the dataset, client environment, and temporary files.
|
| 90 |
|
| 91 |
## Recorded run states
|