MM-IssueLocBench / examples /load_dataset.py
Jasaxion's picture
Initial upload: MM-IssueLoc Bench v1.0 (canonical + function_level parquet, VCE+ preview, repo downloader)
41e2b8e verified
Raw
History Blame
6.84 kB
"""
The dataset ships two Parquet configs under `data/`:
canonical.parquet 652 rows β€” full benchmark; drives file-level eval
function_level.parquet 343 rows β€” subset where edit_functions is non-empty
This script walks through both evaluation granularities (file-level and
function-level), decodes an image, reads a 20-row VCE+ preview from
`examples/preview/vce_plus_preview.jsonl`, and shows how to materialize
the underlying GitHub repository snapshot required at scoring time.
Run:
python3 examples/load_dataset.py
"""
from __future__ import annotations
import json
from pathlib import Path
from datasets import Dataset
ROOT = Path(__file__).resolve().parent.parent
DATA = ROOT / "data"
ds = Dataset.from_parquet(str(DATA / "canonical.parquet"))
ds_fn = Dataset.from_parquet(str(DATA / "function_level.parquet"))
print(f"canonical: {len(ds):4d} rows (file-level eval β€” all edit_files non-empty)")
print(f"function_level: {len(ds_fn):4d} rows (function-level eval β€” strict subset)")
# ===========================================================================
# 1. File-level evaluation
# ===========================================================================
# Input : issue_title + issue_body + images + (repo snapshot at base_commit)
# Output: ranked list of files to edit
# Gold : row["edit_files"] β€” always non-empty in the canonical config.
# ===========================================================================
row = ds[0]
print(f"\n── file-level example ──────────────────────────────────────────────")
print(f" instance_id : {row['instance_id']}")
print(f" repo @ commit: {row['repo_full_name']} @ {row['base_commit'][:12]}")
print(f" language : {row['language']} ({row['language_category']})")
print(f" difficulty : {row['difficulty']} (changed_files={row['changed_files']})")
print(f" # images : {len(row['images'])}")
print(f" gold files : {row['edit_files']}")
print(f" (your model must rank {row['edit_files'][0]!r} near the top)")
# Decode the first screenshot β€” HF Image feature returns PIL.Image automatically.
img = row["images"][0]
print(f" image[0] : {img.size} px, mode={img.mode}, caption={row['image_alts'][0]!r}")
# img.save("first_issue_screenshot.png") # persist for visual inspection
# ===========================================================================
# 2. Function-level evaluation
# ===========================================================================
# Input : same as file-level
# Output: ranked list of `path/to/file:function` identifiers
# Gold : row["edit_functions"] (always non-empty in the function_level config)
# Caveat: exclude row["added_functions"] when scoring β€” those functions do
# not exist at base_commit, so they cannot be retrieved from the
# repository tree.
# ===========================================================================
fn_row = ds_fn[0]
print(f"\n── function-level example ──────────────────────────────────────────")
print(f" instance_id : {fn_row['instance_id']}")
print(f" edit_functions : {fn_row['edit_functions']}")
print(f" added_functions: {fn_row['added_functions']} ← exclude from scoring")
print(f" supports_function_level: {fn_row['supports_function_level']}")
# Pick any instance that is ONLY file-level (no function annotation) to
# illustrate the filter that separates the two configs.
file_only = next(r for r in ds if not r["supports_function_level"] or not r["edit_functions"])
print(f"\n (a file-level-only instance in canonical but NOT in function_level)")
print(f" instance_id : {file_only['instance_id']}")
print(f" supports_function_level : {file_only['supports_function_level']}")
print(f" edit_functions (empty here!) : {file_only['edit_functions']}")
# ===========================================================================
# 3. VCE+ preview (illustrative; not a shipped config)
# ===========================================================================
# VCE+ is a 7-dimensional structured extraction per screenshot (OCR, error
# signal, UI elements, user action, code hints, visual saliency, confidence)
# produced by a multimodal LLM. The FULL cache is a rerunnable byproduct and
# is NOT shipped in data/. A 20-row preview aligned by instance_id with
# examples/preview/preview.jsonl ships at examples/preview/vce_plus_preview.jsonl
# so the schema stays discoverable without running the extractor.
# ===========================================================================
vce_preview_path = ROOT / "examples" / "preview" / "vce_plus_preview.jsonl"
vce_rows = [json.loads(l) for l in vce_preview_path.open()]
print(f"\n── VCE+ preview (20 records) ──────────────────────────────────────")
print(f" path: {vce_preview_path.relative_to(ROOT)} ({len(vce_rows)} rows)")
vce_row = vce_rows[0]
print(f" sample instance_id: {vce_row['instance_id']} (category={vce_row['category']})")
for i, rec in enumerate(vce_row["records"]):
err = rec["error_signal"] or {}
ch = rec["code_hint"] or {}
print(
f" record[{i}]: conf={rec.get('confidence', 0.0):.2f} "
f"error={err.get('kind', '')}/{err.get('type', '')!r} "
f"ui_elements={(rec.get('ui_elements') or [])[:3]} "
f"frameworks={ch.get('frameworks', [])}"
)
# ===========================================================================
# 4. Repository snapshots
# ===========================================================================
# Scoring requires the repository tree at each row's base_commit. Repos are
# not embedded in the Parquet (they would total several GB and carry
# heterogeneous upstream licenses). Use commit_cache.json + the bundled
# script to fetch them:
#
# export GITHUB_TOKEN=ghp_xxx
# python3 scripts/download_repos.py # fetch all 653 repos
# python3 scripts/download_repos.py --only 5 # dry-run slice
# ===========================================================================
cache = json.loads((ROOT / "commit_cache.json").read_text())
entry = cache[row["instance_id"]]
print(f"\n── repo snapshot for the file-level example ─────────────────────────")
print(f" expected directory: repos/{entry['dir_name']}/")
print(f" equivalent to :")
print(f" git clone https://github.com/{entry['repo']} {entry['dir_name']}")
print(f" git -C {entry['dir_name']} checkout {entry['sha']}")
print(f" (or run: python3 scripts/download_repos.py β€” see scripts/download_repos.py)")