Datasets:

Jasaxion
/

MM-IssueLocBench

@@ -31,200 +31,82 @@ configs:
 # MM-IssueLoc Bench
 > **MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization**
-A repository-level, multimodal **issue → source-code localization** benchmark. Given a GitHub issue (title + body + screenshots / diagrams) and a target repository, the task is to retrieve the **file(s)** (and **function(s)**) that must be edited to resolve the issue. This is the first benchmark that pairs real multimodal issues with repository-scoped gold edits at two granularities.
-| # | Task | Input | Output | # Instances |
-|---|---|---|---|---|
-| 1 | File-level localization | Issue text + images + repository tree / contents | Ranked list of files to edit | 652 |
-| 2 | Function-level localization | Issue text + images + repository | Ranked list of `file:function` identifiers | 343 |
----
-## Dataset at a glance
-- **652 canonical instances** (450 human-annotated + 202 AI-augmented).
-- **23 programming languages** — TypeScript 151, Python 126, JavaScript 120, C++ 45, Java 44, Go 33, C# 33, Rust 27, C 21, PHP 20, remainder < 15 each.
 - **7 image categories** — `ui_screenshot` 177, `behavior_demo` 99, `error_message` 92, `rendering_result` 85, `code_screenshot` 84, `log_output` 66, `data_visualization` 50.
-- **3 difficulty buckets** (by `changed_files`) — easy 214 (1 file), medium 263 (2–3 files), hard 176 (≥ 4 files).
-- **1050 issue screenshots embedded** in the canonical config (bytes, no external paths required).
-- **Two-granularity gold labels** — `edit_files` (file level, 652/652 non-empty) and `edit_functions` (function level, 343/343 in the `function_level` config).
----
-## Repository
-Start at `examples/preview/` for a zero-install skim of what an instance looks like; use `examples/load_dataset.py` for the full Parquet-backed workflow, and `scripts/download_repos.py` once you need repository snapshots at `base_commit`.
----
-## Configs
-Two configs are shipped as separate Parquet files; every config contains a single `test` split (this is a pure evaluation benchmark).
-### `canonical` — 652 rows
-The full benchmark. Every instance has non-empty `edit_files`, so this config serves both the **file-level evaluation** directly and is the superset from which `function_level` is derived.
 ```python
 from datasets import load_dataset
-ds = load_dataset("<hub-id>/MM-IssueLoc-Bench", name="canonical", split="test")
 row = ds[0]
-row["images"][0]        # PIL.Image — first screenshot attached to the issue
-row["edit_files"]       # list[str] — gold-standard files to edit
-```
-### `function_level` — 343 rows
-Subset of `canonical` where `supports_function_level == True` **and** `edit_functions` is non-empty. Identifiers follow the `path/to/file.py:Class.method` / `path/to/file.py:top_level_fn` convention. Use this config for function-level metrics.
-```python
-ds = load_dataset("<hub-id>/MM-IssueLoc-Bench", name="function_level", split="test")
-row = ds[0]
-row["edit_functions"]   # list[str] — gold-standard functions to edit
-row["added_functions"]  # list[str] — functions created by the patch (exclude at eval time)
 ```
-## Schema
-### `canonical` & `function_level`
 | Field | Type | Description |
 |---|---|---|
-| `instance_id` | `string` | Unique key, format `<owner>__<repo>__<issue_id>`. |
 | `annotation_by` | `string` | `"human"` or `"ai"`. |
 | `repo_full_name` | `string` | GitHub `owner/repo`. |
-| `repo_language` | `string` | Primary language reported by GitHub. |
-| `language` | `string` | Primary language of the actually edited files (lowercase). |
-| `language_category` | `string` | Coarse bucket: `frontend` / `backend` / `systems` / `data_science` / … |
-| `base_commit` | `string` | PR base commit SHA — checkout the repo here before evaluation. |
-| `issue_title` | `string` | Issue title. |
-| `issue_body` | `string` | Issue body in original markdown (images are referenced as markdown; see `images`). |
-| `images` | `Sequence[Image]` | Issue screenshots, decoded as `PIL.Image` when loaded with 🤗 `datasets`. |
-| `image_paths` | `Sequence[string]` | Original filenames (`issue_<issue_id>_<idx>.<ext>`), aligned with `images`. |
-| `image_alts` | `Sequence[string]` | Markdown `alt` text per image. |
-| `image_sources` | `Sequence[string]` | Per-image provenance tag (`body`, `comment`, …). |
-| `image_category` | `string` | One of the 7 image categories. |
-| `relevance_score` | `int32` | Annotator score for image–issue relevance. |
-| `difficulty` | `string` | `easy` / `medium` / `hard` — bucketed by `changed_files` (see below). |
-| `diff` | `string` | Full unified diff of the resolving PR. |
-| `diff_files` | `Sequence[string]` | Files touched by `diff`. |
-| `diff_status` | `string` | Diff-extraction status flag (e.g. `"ok"`). |
-| `edit_files` | `Sequence[string]` | **Gold standard** for file-level evaluation. |
-| `edit_functions` | `Sequence[string]` | **Gold standard** for function-level evaluation (may be empty in `canonical`). |
 | `added_functions` | `Sequence[string]` | Functions introduced by the patch — exclude at function-level eval time. |
-| `supports_function_level` | `bool` | Whether function-level evaluation is supported for this instance. |
-| `additions` | `int32` | Lines added in the diff. |
-| `deletions` | `int32` | Lines deleted in the diff. |
-| `changed_files` | `int32` | Number of files touched. |
-| `patch_count` | `int32` | Number of patches in the PR. |
-| `repo_stars` | `int32` | Star count at collection time. |
-| `repo_license` | `string` | Source repository license (SPDX-style where available). |
-### `vce_plus_cache` (preview only — see `examples/preview/vce_plus_preview.jsonl`)
-| Field | Type | Description |
-|---|---|---|
-| `instance_id` | `string` | Foreign key to `canonical.instance_id`. |
-| `category` | `string` | Image category (redundant with `canonical.image_category`). |
-| `records` | `List[Struct]` | One entry per screenshot; see below. |
-Per-record struct:
-| Field | Type | Description |
-|---|---|---|
-| `ocr_text` | `string` | OCR result. |
-| `error_signal` | `struct` | `{kind, type, message, stack_hint[]}` — extracted error semantics. |
-| `ui_elements` | `Sequence[string]` | Detected UI widgets. |
-| `user_action` | `string` | Inferred user intent / action. |
-| `code_hint` | `struct` | `{function_names[], file_patterns[], frameworks[]}` ,code-side hooks. |
-| `visual_saliency` | `string` | Free-text description of the salient region. |
-| `confidence` | `float32` | Extractor confidence ∈ [0, 1]. |
-| `notes` | `string` | Extractor-side remarks. |
-### `difficulty` bucketing
-Bucketed on `changed_files` of the resolving PR. Rule is 100 % consistent with the 402 human-labelled difficulty fields in the original annotation pass (audited); the remaining 251 AI-augmented rows were back-filled with the same rule.
-| Bucket | Rule | Count | Median `additions` | Median `deletions` |
-|---|---|---:|---:|---:|
-| `easy` | `changed_files == 1` | 214 | 6 | 2 |
-| `medium` | `changed_files ∈ {2, 3}` | 263 | 28 | 6 |
-| `hard` | `changed_files ≥ 4` | 176 | 59 | 16 |
----
-## Evaluation protocol
-### Repository snapshot
-Every instance is evaluated against the repository state at its `base_commit`. The companion file `commit_cache.json` (at the root of this dataset) maps `instance_id → {repo, sha, dir_name}` and is consumed by the bundled download script.
-**Recommended — use the bundled downloader.** Handles rate-limit-aware concurrent fetching, atomic extraction, and resumable retries out of the box:
 ```bash
 export GITHUB_TOKEN=ghp_xxx
-python3 scripts/download_repos.py
-python3 scripts/download_repos.py --workers 8  # faster with more parallelism
-python3 scripts/download_repos.py --retry-failed  # after an interruption
 ```
-The script uses the GitHub REST tarball API (`GET /repos/{owner}/{repo}/tarball/{sha}`), so a **personal access token with the `public_repo` scope is strongly recommended** (generate at <https://github.com/settings/tokens>; no elevated permissions needed for public repositories). Without a token, you are throttled to 60 requests/hour, which makes the full 653-repo download take ~11 hours instead of ~10 minutes.
-On success you get:
-```
-repos/
-├── fbaligand__kibana-enhanced-table__569e9d9ee2fb/
-├── presscustomizr__hueman__0f3aeb3c5b2a/
-└── ... (653 total; naming = <owner>__<repo>__<sha12>, no .git)
-```
-Failed downloads (archived / deleted / moved repos) are logged to `download_failures.json` for targeted re-runs.
-Repository tarballs are **not** shipped here (they total several GB and individual repos have heterogeneous licenses); users fetch them from GitHub on demand using the script or the cache above.
-### Metrics
-File-level and function-level evaluation are tracked independently using the same family of ranking metrics at `K ∈ {1, 3, 5, 10}
-- **Acc@K** — set-based hit rates.
-- Other metrics avaliable:
-  - **MRR** — Mean Reciprocal Rank of the first hit.
-  - **Recall@K**, **Hit@K** — set-based hit rates.
-  - **MAP@K**, **NDCG@K** — order-sensitive metrics.
-Recommended normalisation at scoring time:
-1. Path normalisation — strip `./`, collapse separators, lowercase drive letters on Windows.
-2. Function ID matching — accept `file:func` ≡ `file:Class.func` (relaxed class-qualifier match).
 ## Intended use & limitations
-**Intended uses.**
-- Benchmarking repository-level issue localization models, with or without multimodal inputs.
-- Ablating the contribution of screenshots vs. text to localization accuracy.
-- Studying how different image categories (UI screenshots vs. error messages vs. log output) help or hurt retrieval.
-**Out-of-scope uses.**
-- Patch generation / code repair — this dataset ships gold edits but is evaluated on localization, not synthesis. The `diff` field is provided for offline analysis but is not a supervised target.
-- Training large-scale multimodal models end-to-end — the instance count (652) is deliberately sized for evaluation, not pre-training.
-- **Commercial use**
-**Known limitations.**
-- Language skew towards web-ecosystem repos (TypeScript / Python / JavaScript together cover ~60 %).
-- Repositories are real open-source projects; code licensing varies per row (`repo_license`).
-## Ethics & privacy
-All issues are drawn from public GitHub repositories. Screenshots occasionally contain user-visible strings from demo data (e.g., placeholder usernames, example URLs). We reviewed the top-frequency tokens in image OCR output and found no personally identifying information beyond public GitHub author handles already visible in `instance_id` / `repo_full_name`. No private data, no scraped user content outside the public issue tracker, no model-generated synthetic faces or identifiers.
----
 ## Citation

 # MM-IssueLoc Bench
 > **MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization**
+> [📄 arXiv](https://arxiv.org/abs/2607.15205) · [💻 Evaluation toolkit](https://github.com/Jasaxion/MM-IssueLoc-Bench)
+Repository-level, multimodal **issue → source-code localization**. Given a GitHub issue (title + body + screenshots) and a target repository, retrieve the **file(s)** and **function(s)** that must be edited to resolve it — with visual evidence treated as an explicit input, decoupled from patch synthesis.
+| Config | Task | Output | Instances |
+|---|---|---|---|
+| `canonical` | File-level localization | Ranked files to edit | 652 |
+| `function_level` | Function-level localization | Ranked `file:function` ids | 343 |
+Both configs ship a single `test` split (pure evaluation benchmark).
+## At a glance
+- **652 instances** (450 human-annotated + 202 AI-augmented), **1050 screenshots** embedded as bytes.
+- **23 languages** — TypeScript 151, Python 126, JavaScript 120, C++ 45, Java 44, Go 33, C# 33, Rust 27, C 21, PHP 20, rest < 15.
 - **7 image categories** — `ui_screenshot` 177, `behavior_demo` 99, `error_message` 92, `rendering_result` 85, `code_screenshot` 84, `log_output` 66, `data_visualization` 50.
+- **3 difficulty buckets** (by `changed_files`) — easy 214 (1), medium 263 (2–3), hard 176 (≥4).
+- **Two-granularity gold** — `edit_files` (all 652) and `edit_functions` (343, in `function_level`).
+## Usage
 ```python
 from datasets import load_dataset
+ds = load_dataset("Jasaxion/MM-IssueLocBench", name="canonical", split="test")
 row = ds[0]
+row["images"][0]        # PIL.Image — first issue screenshot
+row["edit_files"]       # list[str] — gold files to edit
+fn = load_dataset("Jasaxion/MM-IssueLocBench", name="function_level", split="test")
+fn[0]["edit_functions"] # list[str] — gold `path/to/file.py:Class.method` ids
 ```
+`function_level` is the subset of `canonical` where `supports_function_level` is
+true and `edit_functions` is non-empty. For scoring, the
+[evaluation toolkit](https://github.com/Jasaxion/MM-IssueLoc-Bench) provides the loader, metrics (Acc@K, MRR, Recall@K, Hit@K, MAP@K, NDCG@K), and CLIs.
+## Schema
 | Field | Type | Description |
 |---|---|---|
+| `instance_id` | `string` | Unique key `<owner>__<repo>__<issue_id>`. |
 | `annotation_by` | `string` | `"human"` or `"ai"`. |
 | `repo_full_name` | `string` | GitHub `owner/repo`. |
+| `repo_language` / `language` | `string` | GitHub primary language / language of edited files. |
+| `language_category` | `string` | `frontend` / `backend` / `systems` / `data_science` / … |
+| `base_commit` | `string` | PR base commit SHA — check out the repo here before evaluation. |
+| `issue_title` / `issue_body` | `string` | Issue text (body in original markdown). |
+| `images` | `Sequence[Image]` | Screenshots, decoded as `PIL.Image`. |
+| `image_paths` / `image_alts` / `image_sources` | `Sequence[string]` | Per-image filename, alt text, provenance (`body` / `comment` / …), aligned with `images`. |
+| `image_category` | `string` | One of the 7 categories. |
+| `relevance_score` | `int32` | Annotator image–issue relevance score. |
+| `difficulty` | `string` | `easy` / `medium` / `hard`, bucketed by `changed_files`. |
+| `diff` | `string` | Full unified diff of the resolving PR (offline analysis only). |
+| `diff_files` / `diff_status` | `Sequence[string]` / `string` | Files touched by `diff` / extraction status. |
+| `edit_files` | `Sequence[string]` | **Gold** for file-level evaluation. |
+| `edit_functions` | `Sequence[string]` | **Gold** for function-level evaluation (may be empty in `canonical`). |
 | `added_functions` | `Sequence[string]` | Functions introduced by the patch — exclude at function-level eval time. |
+| `supports_function_level` | `bool` | Whether function-level evaluation applies. |
+| `additions` / `deletions` / `changed_files` / `patch_count` | `int32` | Diff size statistics. |
+| `repo_stars` / `repo_license` | `int32` / `string` | Repo stars at collection / SPDX license. |
+## Repository snapshots
+Evaluation runs against each repo at its `base_commit`; tarballs are **not** shipped (several GB, heterogeneous licenses). `commit_cache.json` maps `instance_id → {repo, sha, dir_name}`, and `scripts/download_repos.py` fetches them from the GitHub tarball API (a `public_repo`-scoped `GITHUB_TOKEN` is strongly recommended to avoid the 60 req/hour limit):
 ```bash
 export GITHUB_TOKEN=ghp_xxx
+python3 scripts/download_repos.py --workers 8   # --retry-failed to resume
 ```
+Snapshots land under `repos/<owner>__<repo>__<sha12>/`; failures are logged to `download_failures.json`. See `examples/preview/` for a zero-install skim of the data and `examples/load_dataset.py` for the full workflow.
 ## Intended use & limitations
+For benchmarking repository-level issue localization. **Out of scope:** patch generation, end-to-end training (652 instances is an evaluation set), and commercial use (CC BY-NC 4.0). Content skews toward web-ecosystem repos (TS/Py/JS ≈ 60%); per-row code licenses vary (`repo_license`). All data is from public GitHub issues and contains no PII beyond already-public author handles.
 ## Citation