--- pretty_name: Process-Aware Code Review language: - en tags: - code-review - coding-agent - trajectory - deepswe - atif size_categories: - n<1K configs: - config_name: all default: true data_files: - split: train path: data/*/train.jsonl - config_name: gpt55 data_files: - split: train path: data/gpt55/train.jsonl - config_name: gpt56 data_files: - split: train path: data/gpt56/train.jsonl - config_name: gemini data_files: - split: train path: data/gemini/train.jsonl ---

Diffs Are Cheap, Show Me the Trajectory

Process-Aware Code Review
Coding trajectories, synthetic reference comments, and implementation patches.

Quick start   /   Data overview   /   Data fields   /   Code and experiments

## What is included **113 DeepSWE tasks, three coding models, 339 complete coding trajectories, and 882 synthetic reference comments.** Each record pairs a coding trajectory with its implementation patch and reference review comments. | Core data | Contents | Field | | :--- | :--- | :--- | | Coding trajectories | Complete ATIF-v1.7 development records, including agent messages and tool interactions | `traj` | | Reference comments | Synthetic reference findings with original descriptions, evidence, and metadata | `review_comments` | | Implementation patches | Coding-agent changes relative to the upstream base commit | `patch` | | Task context | Task description, upstream repository, programming language, and base commit | `task`, `repository`, `language`, `base_commit` | | Reference solutions | Held-out reference diffs, separate from coding-agent implementations | `reference_patch` | | Verification outcomes | Recorded pass, failure, or harness timeout | `verifier_status` | All **339 coding-agent trajectories** are complete, not summaries or truncated exports. Each coding model implements the same 113 tasks; these are **339 model-task records, not 339 distinct tasks**. **Not included:** standalone benchmark task definitions, task Dockerfiles, hidden test patches, repository snapshots, reviewer execution traces, or the full experimental reproduction code. See the [source repository](https://github.com/Jacksadventure/review_in_process) for the broader experimental artifact. ## Find the coding trajectories | Configuration | Coding model | Records | Reference comments | Data | | :--- | :--- | ---: | ---: | :--- | | `gpt55` | GPT-5.5 | 113 | 355 | [JSONL](data/gpt55/train.jsonl) | | `gpt56` | GPT-5.6 Sol | 113 | 185 | [JSONL](data/gpt56/train.jsonl) | | `gemini` | Gemini 3.8 Flash | 113 | 342 | [JSONL](data/gemini/train.jsonl) | | `all` | All three models | 339 | 882 | All files above | The default `all` configuration combines the three coding cohorts. Model configurations are not independent train/test splits. The single `train` split is a storage convention, not a prescribed evaluation protocol. ## Quick start ```sh pip install datasets ``` ```python import json from datasets import load_dataset data = load_dataset("ZijiantheJack/ProcessAwareCodeReview", "all", split="train") example = data[0] trajectory = json.loads(example["traj"]) reference_review = json.loads(example["review_comments"]) comments = reference_review["reviews"] implementation_patch = example["patch"] ``` Select `gpt55`, `gpt56`, or `gemini` instead of `all` to load one coding model. Use `instance_id` to align the three implementations of the same task. ## Data fields | Field | Contents | | :--- | :--- | | `id` | Unique `coding_id/instance_id` key | | `instance_id` | Task identifier shared across the three coding models | | `coding_id`, `coding_model` | Coding configuration and model name | | `repository`, `base_commit`, `language` | Upstream project context | | `traj` | Complete original ATIF-v1.7 JSON text; parse with `json.loads` | | `review_comments` | Complete original reference-review JSON text, including `reviews` and validation metadata | | `patch` | Original coding-agent implementation diff | | `reference_patch` | Separate reference solution diff | | `task` | Task description supplied to the coding agent | | `reference_count` | Number of comments in the reference record | | `verifier_status` | Recorded official outcome: `passed`, `failed`, or `harness_timeout` | JSON-valued fields are UTF-8 strings so heterogeneous native ATIF steps and reference schemas load without flattening, truncation, or schema inference loss. After JSONL decoding, these strings and the patch/task fields preserve the original source text exactly. Each reference comment retains its original ID, title, description, expected behavior, severity, file, region, failure scenario, supporting evidence, and confidence. Empty comment lists are retained: **2 GPT-5.5, 21 GPT-5.6 Sol, and 3 Gemini records** have no reference comments. No implementation patch is empty. ## Synthetic reference comments Reference comments are **model-generated reference labels, not human reviews** or outputs from the different review methods in the broader experiment. Reference generation uses task text and implementation/reference patches, without consuming the coding trajectory. Original metadata are retained, including `reference_kind`, `validation_status`, and `deduplication_status`. Records include `synthetic_unvalidated_signal_closure` with validation and deduplication skipped; a reference label is not a guarantee of a verified defect. The original comment `severity` is not a newly inferred failure-linked label. Trajectories contain original agent messages, tool inputs and outputs, code excerpts, generated tests, and recorded local paths. Model-produced content and findings can be incorrect. ## Verification outcomes | Coding model | Passed | Failed | Harness timeout | | :--- | ---: | ---: | ---: | | GPT-5.5 | 62 | 51 | 0 | | GPT-5.6 Sol | 82 | 31 | 0 | | Gemini 3.8 Flash | 84 | 28 | 1 | | **Total** | **228** | **110** | **1** | These are recorded outcomes, not newly executed verification. A timeout is not silently treated as a pass or failure. ## Layout ```text data/ gpt55/train.jsonl gpt56/train.jsonl gemini/train.jsonl assets/ logo.svg provenance/ LICENSES/ artifact/origin.json docs/PROVENANCE.md export-report.json source-paths.json SHA256SUMS.json ``` ## Rights and attribution The source artifact does not grant a blanket license for all included content. DeepSWE's Apache-2.0 notice covers its original contributions and does not relicense upstream code. Task patches remain subject to their upstream projects' terms; the source framework's ISC declaration does not relicense benchmark data. Retained notices: [DeepSWE Apache-2.0](provenance/LICENSES/DeepSWE-Apache-2.0.txt) and [upstream attribution](provenance/LICENSES/UPSTREAM-PROVENANCE.md). Consult those records and the relevant upstream terms before redistribution or reuse. This dataset does not assert new licensing rights.