Refresh dataset homepage using original artifact README style
Browse files- README.md +152 -78
- SHA256SUMS.json +6 -2
- assets/logo.svg +37 -0
README.md
CHANGED
|
@@ -30,45 +30,87 @@ configs:
|
|
| 30 |
path: data/gemini/train.jsonl
|
| 31 |
---
|
| 32 |
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
```python
|
| 50 |
import json
|
| 51 |
from datasets import load_dataset
|
| 52 |
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
|
|
|
| 59 |
```
|
| 60 |
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
does not establish a train/test partition.
|
| 64 |
|
| 65 |
-
##
|
| 66 |
|
| 67 |
| Field | Contents |
|
| 68 |
-
| --- | --- |
|
| 69 |
| `id` | Unique `coding_id/instance_id` key |
|
| 70 |
-
| `instance_id` | Task identifier
|
| 71 |
-
| `coding_id`, `coding_model` | Coding
|
| 72 |
| `repository`, `base_commit`, `language` | Upstream project context |
|
| 73 |
| `traj` | Complete original ATIF-v1.7 JSON text; parse with `json.loads` |
|
| 74 |
| `review_comments` | Complete original reference-review JSON text, including `reviews` and validation metadata |
|
|
@@ -76,66 +118,98 @@ does not establish a train/test partition.
|
|
| 76 |
| `reference_patch` | Separate reference solution diff |
|
| 77 |
| `task` | Task description supplied to the coding agent |
|
| 78 |
| `reference_count` | Number of comments in the reference record |
|
| 79 |
-
| `verifier_status` | Recorded official
|
| 80 |
|
| 81 |
JSON-valued fields are UTF-8 strings so heterogeneous native ATIF steps and
|
| 82 |
-
|
| 83 |
-
After JSONL decoding, these strings and patch/task fields preserve the
|
| 84 |
-
source text exactly.
|
| 85 |
-
|
| 86 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 87 |
|
| 88 |
-
##
|
| 89 |
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
present. A reference label is not a guarantee of a verified defect. Source
|
| 95 |
-
reference generation did not consume the coding trajectory.
|
| 96 |
|
| 97 |
-
The
|
| 98 |
-
|
|
|
|
|
|
|
| 99 |
|
| 100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
each of the 1,586 payload/provenance source files was verified against the
|
| 106 |
-
requested GitHub repository's Git blob IDs. All exported record fields and
|
| 107 |
-
all 339 model-task keys match that commit's Stage 2 index. The live index adds
|
| 108 |
-
an alternative `standalone_references` path for Gemini; this export consistently
|
| 109 |
-
uses the original `references` records for all three models.
|
| 110 |
-
|
| 111 |
-
`data/stage2.json` in that archive binds each model-task record to its coding
|
| 112 |
-
trajectory, implementation patch, and reference comments. Every exported source
|
| 113 |
-
file was checked against the archive's embedded SHA-256 integrity manifest.
|
| 114 |
-
Relative symlinks were resolved within the archive and both their link text
|
| 115 |
-
and resolved content hashes were checked. This checks internal consistency;
|
| 116 |
-
the exported payloads were additionally verified against the GitHub commit above.
|
| 117 |
-
|
| 118 |
-
- `SHA256SUMS.json`: file sizes and SHA-256 hashes for this export.
|
| 119 |
-
- `provenance/export-report.json`: archive digest, cohort counts, zero-comment
|
| 120 |
-
cases, verification outcomes, and verified source-file hashes.
|
| 121 |
-
- `provenance/source-paths.json`: original source paths for each record.
|
| 122 |
-
- `provenance/source-stage2-index.json`: the verified GitHub commit's Stage 2 index.
|
| 123 |
-
- `provenance/artifact/origin.json`: original artifact provenance.
|
| 124 |
-
- `provenance/docs/PROVENANCE.md`: original scope and provenance statement.
|
| 125 |
-
|
| 126 |
-
Benchmark task definitions, hidden test patches, Dockerfiles, repository
|
| 127 |
-
snapshots, reviewer debug logs, and live credentials are not part of this export.
|
| 128 |
|
| 129 |
## Rights and attribution
|
| 130 |
|
| 131 |
The source artifact does not grant a blanket license for all included content.
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
- [DeepSWE Apache-2.0 notice](provenance/LICENSES/DeepSWE-Apache-2.0.txt)
|
| 137 |
-
- [Upstream attribution and rights](provenance/LICENSES/UPSTREAM-PROVENANCE.md)
|
| 138 |
|
|
|
|
|
|
|
|
|
|
| 139 |
Consult those records and the relevant upstream terms before redistribution
|
| 140 |
-
or reuse. This dataset does not
|
| 141 |
-
comments, a new benchmark split, or new licensing rights.
|
|
|
|
| 30 |
path: data/gemini/train.jsonl
|
| 31 |
---
|
| 32 |
|
| 33 |
+
<p align="center">
|
| 34 |
+
<img src="assets/logo.svg" alt="Diffs Are Cheap, Show Me the Trajectory" width="100%">
|
| 35 |
+
</p>
|
| 36 |
+
<p align="center">
|
| 37 |
+
<strong>Process-Aware Code Review</strong><br>
|
| 38 |
+
Coding trajectories, synthetic reference comments, and implementation patches.
|
| 39 |
+
</p>
|
| 40 |
+
<p align="center">
|
| 41 |
+
<a href="#quick-start">Quick start</a> /
|
| 42 |
+
<a href="#what-is-included">Data overview</a> /
|
| 43 |
+
<a href="#data-fields">Data fields</a> /
|
| 44 |
+
<a href="https://github.com/Jacksadventure/review_in_process">Code and experiments</a>
|
| 45 |
+
</p>
|
| 46 |
+
|
| 47 |
+
## What is included
|
| 48 |
+
|
| 49 |
+
**113 DeepSWE tasks, three coding models, 339 complete coding trajectories,
|
| 50 |
+
and 882 synthetic reference comments.** Each record pairs a coding trajectory
|
| 51 |
+
with its implementation patch and reference review comments.
|
| 52 |
+
|
| 53 |
+
| Core data | Contents | Field |
|
| 54 |
+
| :--- | :--- | :--- |
|
| 55 |
+
| Coding trajectories | Complete ATIF-v1.7 development records, including agent messages and tool interactions | `traj` |
|
| 56 |
+
| Reference comments | Synthetic reference findings with original descriptions, evidence, and metadata | `review_comments` |
|
| 57 |
+
| Implementation patches | Coding-agent changes relative to the upstream base commit | `patch` |
|
| 58 |
+
| Task context | Task description, upstream repository, programming language, and base commit | `task`, `repository`, `language`, `base_commit` |
|
| 59 |
+
| Reference solutions | Held-out reference diffs, separate from coding-agent implementations | `reference_patch` |
|
| 60 |
+
| Verification outcomes | Recorded pass, failure, or harness timeout | `verifier_status` |
|
| 61 |
+
|
| 62 |
+
All **339 coding-agent trajectories** are complete, not summaries or truncated
|
| 63 |
+
exports. Each coding model implements the same 113 tasks; these are **339
|
| 64 |
+
model-task records, not 339 distinct tasks**.
|
| 65 |
+
|
| 66 |
+
**Not included:** standalone benchmark task definitions, task Dockerfiles,
|
| 67 |
+
hidden test patches, repository snapshots, reviewer execution traces, or the
|
| 68 |
+
full experimental reproduction code. See the
|
| 69 |
+
[source repository](https://github.com/Jacksadventure/review_in_process)
|
| 70 |
+
for the broader experimental artifact.
|
| 71 |
+
|
| 72 |
+
## Find the coding trajectories
|
| 73 |
+
|
| 74 |
+
| Configuration | Coding model | Records | Reference comments | Data |
|
| 75 |
+
| :--- | :--- | ---: | ---: | :--- |
|
| 76 |
+
| `gpt55` | GPT-5.5 | 113 | 355 | [JSONL](data/gpt55/train.jsonl) |
|
| 77 |
+
| `gpt56` | GPT-5.6 Sol | 113 | 185 | [JSONL](data/gpt56/train.jsonl) |
|
| 78 |
+
| `gemini` | Gemini 3.8 Flash | 113 | 342 | [JSONL](data/gemini/train.jsonl) |
|
| 79 |
+
| `all` | All three models | 339 | 882 | All files above |
|
| 80 |
+
|
| 81 |
+
The default `all` configuration combines the three coding cohorts. Model
|
| 82 |
+
configurations are not independent train/test splits. The single `train` split
|
| 83 |
+
is a storage convention, not a prescribed evaluation protocol.
|
| 84 |
+
|
| 85 |
+
## Quick start
|
| 86 |
+
|
| 87 |
+
```sh
|
| 88 |
+
pip install datasets
|
| 89 |
+
```
|
| 90 |
|
| 91 |
```python
|
| 92 |
import json
|
| 93 |
from datasets import load_dataset
|
| 94 |
|
| 95 |
+
data = load_dataset("ZijiantheJack/ProcessAwareCodeReview", "all", split="train")
|
| 96 |
+
example = data[0]
|
| 97 |
+
|
| 98 |
+
trajectory = json.loads(example["traj"])
|
| 99 |
+
reference_review = json.loads(example["review_comments"])
|
| 100 |
+
comments = reference_review["reviews"]
|
| 101 |
+
implementation_patch = example["patch"]
|
| 102 |
```
|
| 103 |
|
| 104 |
+
Select `gpt55`, `gpt56`, or `gemini` instead of `all` to load one coding model.
|
| 105 |
+
Use `instance_id` to align the three implementations of the same task.
|
|
|
|
| 106 |
|
| 107 |
+
## Data fields
|
| 108 |
|
| 109 |
| Field | Contents |
|
| 110 |
+
| :--- | :--- |
|
| 111 |
| `id` | Unique `coding_id/instance_id` key |
|
| 112 |
+
| `instance_id` | Task identifier shared across the three coding models |
|
| 113 |
+
| `coding_id`, `coding_model` | Coding configuration and model name |
|
| 114 |
| `repository`, `base_commit`, `language` | Upstream project context |
|
| 115 |
| `traj` | Complete original ATIF-v1.7 JSON text; parse with `json.loads` |
|
| 116 |
| `review_comments` | Complete original reference-review JSON text, including `reviews` and validation metadata |
|
|
|
|
| 118 |
| `reference_patch` | Separate reference solution diff |
|
| 119 |
| `task` | Task description supplied to the coding agent |
|
| 120 |
| `reference_count` | Number of comments in the reference record |
|
| 121 |
+
| `verifier_status` | Recorded official outcome: `passed`, `failed`, or `harness_timeout` |
|
| 122 |
|
| 123 |
JSON-valued fields are UTF-8 strings so heterogeneous native ATIF steps and
|
| 124 |
+
reference schemas load without flattening, truncation, or schema inference loss.
|
| 125 |
+
After JSONL decoding, these strings and the patch/task fields preserve the
|
| 126 |
+
original source text exactly.
|
| 127 |
+
|
| 128 |
+
Each reference comment retains its original ID, title, description, expected
|
| 129 |
+
behavior, severity, file, region, failure scenario, supporting evidence, and
|
| 130 |
+
confidence. Empty comment lists are retained: **2 GPT-5.5, 21 GPT-5.6 Sol, and
|
| 131 |
+
3 Gemini records** have no reference comments. No implementation patch is empty.
|
| 132 |
+
|
| 133 |
+
## Synthetic reference comments
|
| 134 |
+
|
| 135 |
+
Reference comments are **model-generated reference labels, not human reviews**
|
| 136 |
+
or outputs from the different review methods in the broader experiment.
|
| 137 |
+
Reference generation uses task text and implementation/reference patches,
|
| 138 |
+
without consuming the coding trajectory.
|
| 139 |
+
|
| 140 |
+
Original metadata are retained, including `reference_kind`,
|
| 141 |
+
`validation_status`, and `deduplication_status`. Records include
|
| 142 |
+
`synthetic_unvalidated_signal_closure` with validation and deduplication
|
| 143 |
+
skipped; a reference label is not a guarantee of a verified defect.
|
| 144 |
+
The original comment `severity` is not a newly inferred failure-linked label.
|
| 145 |
+
|
| 146 |
+
Trajectories contain original agent messages, tool inputs and outputs, code
|
| 147 |
+
excerpts, generated tests, and recorded local paths. Model-produced content and
|
| 148 |
+
findings can be incorrect.
|
| 149 |
+
|
| 150 |
+
## Verification outcomes
|
| 151 |
+
|
| 152 |
+
| Coding model | Passed | Failed | Harness timeout |
|
| 153 |
+
| :--- | ---: | ---: | ---: |
|
| 154 |
+
| GPT-5.5 | 62 | 51 | 0 |
|
| 155 |
+
| GPT-5.6 Sol | 82 | 31 | 0 |
|
| 156 |
+
| Gemini 3.8 Flash | 84 | 28 | 1 |
|
| 157 |
+
| **Total** | **228** | **110** | **1** |
|
| 158 |
+
|
| 159 |
+
These are recorded outcomes, not newly executed verification. A timeout is not
|
| 160 |
+
silently treated as a pass or failure.
|
| 161 |
+
|
| 162 |
+
## Layout
|
| 163 |
+
|
| 164 |
+
```text
|
| 165 |
+
data/
|
| 166 |
+
gpt55/train.jsonl
|
| 167 |
+
gpt56/train.jsonl
|
| 168 |
+
gemini/train.jsonl
|
| 169 |
+
assets/
|
| 170 |
+
logo.svg
|
| 171 |
+
provenance/
|
| 172 |
+
LICENSES/
|
| 173 |
+
artifact/origin.json
|
| 174 |
+
docs/PROVENANCE.md
|
| 175 |
+
export-report.json
|
| 176 |
+
source-paths.json
|
| 177 |
+
SHA256SUMS.json
|
| 178 |
+
```
|
| 179 |
|
| 180 |
+
## Source and integrity
|
| 181 |
|
| 182 |
+
Source: [Jacksadventure/review_in_process](https://github.com/Jacksadventure/review_in_process).
|
| 183 |
+
Verified commit: `92584bc2d3fd84f83a59398b82a77aeffa81d5b4`.
|
| 184 |
+
All 339 exported trajectory/comment/patch triples were checked byte-for-byte
|
| 185 |
+
against the source repository retrieved via SSH.
|
|
|
|
|
|
|
| 186 |
|
| 187 |
+
The initial export was read from the `review_in_process-1.0.4.zip` artifact;
|
| 188 |
+
all 1,586 payload/provenance source files were additionally verified against the
|
| 189 |
+
requested GitHub commit by Git blob ID. The export consistently uses the original
|
| 190 |
+
`references` records for all three models.
|
| 191 |
|
| 192 |
+
| Record | Purpose |
|
| 193 |
+
| :--- | :--- |
|
| 194 |
+
| [File checksums](SHA256SUMS.json) | Exported file sizes and SHA-256 hashes |
|
| 195 |
+
| [Export report](provenance/export-report.json) | Source archive digest, cohort counts, zero-comment cases, outcomes, and verified source hashes |
|
| 196 |
+
| [Source paths](provenance/source-paths.json) | Original source paths for each record |
|
| 197 |
+
| [Artifact origin](provenance/artifact/origin.json) | Original artifact provenance |
|
| 198 |
+
| [Provenance and scope](provenance/docs/PROVENANCE.md) | Source scope and provenance statement |
|
| 199 |
|
| 200 |
+
The broader source repository contains the review methods, native evaluation,
|
| 201 |
+
analysis, and reproduction tools. Those components are not represented as
|
| 202 |
+
additional records in this dataset.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 203 |
|
| 204 |
## Rights and attribution
|
| 205 |
|
| 206 |
The source artifact does not grant a blanket license for all included content.
|
| 207 |
+
DeepSWE's Apache-2.0 notice covers its original contributions and does not
|
| 208 |
+
relicense upstream code. Task patches remain subject to their upstream projects'
|
| 209 |
+
terms; the source framework's ISC declaration does not relicense benchmark data.
|
|
|
|
|
|
|
|
|
|
| 210 |
|
| 211 |
+
Retained notices:
|
| 212 |
+
[DeepSWE Apache-2.0](provenance/LICENSES/DeepSWE-Apache-2.0.txt) and
|
| 213 |
+
[upstream attribution](provenance/LICENSES/UPSTREAM-PROVENANCE.md).
|
| 214 |
Consult those records and the relevant upstream terms before redistribution
|
| 215 |
+
or reuse. This dataset does not assert new licensing rights.
|
|
|
SHA256SUMS.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
{
|
| 2 |
"README.md": {
|
| 3 |
-
"bytes":
|
| 4 |
-
"sha256": "
|
| 5 |
},
|
| 6 |
"data/gemini/train.jsonl": {
|
| 7 |
"bytes": 108188050,
|
|
@@ -42,5 +42,9 @@
|
|
| 42 |
"provenance/source-stage2-index.json": {
|
| 43 |
"bytes": 564273,
|
| 44 |
"sha256": "6479d35b1a9be66be51335f73c57fc0e08baafd55540d9080d1811109da308ce"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
}
|
| 46 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"README.md": {
|
| 3 |
+
"bytes": 8546,
|
| 4 |
+
"sha256": "18a292df2054a4ad463301ac563fb961c210fd70e768c22fe7de1b94841bc065"
|
| 5 |
},
|
| 6 |
"data/gemini/train.jsonl": {
|
| 7 |
"bytes": 108188050,
|
|
|
|
| 42 |
"provenance/source-stage2-index.json": {
|
| 43 |
"bytes": 564273,
|
| 44 |
"sha256": "6479d35b1a9be66be51335f73c57fc0e08baafd55540d9080d1811109da308ce"
|
| 45 |
+
},
|
| 46 |
+
"assets/logo.svg": {
|
| 47 |
+
"bytes": 3111,
|
| 48 |
+
"sha256": "cfdea6a39d948df4c8d10153b5f1d1941db80f5b310520d652dace0e9ed396a1"
|
| 49 |
}
|
| 50 |
}
|
assets/logo.svg
ADDED
|
|