Datasets:
File size: 3,886 Bytes
6eb86fa | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | #!/usr/bin/env python3
from __future__ import annotations
import argparse
import json
import os
from pathlib import Path
REPO_ID = "Gyubeum/AndroidFlux_RM_Eval"
CONFIGS = ["ui_genie_pairwise", "android_flux_recovery_action_preference"]
def get_parquet(config: str, name: str, local_dir: str | None, token: str | None) -> str:
if local_dir:
return str(Path(local_dir) / config / name)
from huggingface_hub import hf_hub_download
return hf_hub_download(
repo_id=REPO_ID, repo_type="dataset",
filename=f"{config}/{name}", token=token,
)
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--config", default="ui_genie_pairwise", choices=CONFIGS)
ap.add_argument("--subset", default="full", choices=["full", "reduced"])
ap.add_argument("--n", type=int, default=2)
ap.add_argument("--local_dir", default=None,
help="Read the parquet from here instead of the Hub")
ap.add_argument("--image_dir", default="/tmp/rm_eval_images",
help="Where to write screenshots")
args = ap.parse_args()
token = os.environ.get("HF_TOKEN")
import pyarrow.parquet as pq
print("=" * 68)
print(f"1. {args.config}/samples.parquet")
print("=" * 68)
table = pq.read_table(get_parquet(args.config, "samples.parquet", args.local_dir, token))
rows = table.to_pylist()
if args.subset == "reduced":
rows = [r for r in rows if r["in_reduced"]]
print(f" {len(rows)} rows ({args.subset}), {len(table.schema)} columns")
print(f" columns: {', '.join(table.schema.names)}")
rows = rows[: args.n]
for r in rows:
print(f"\n sample_id : {r['sample_id']}")
print(f" goal : {r['goal'][:90]}")
print(f" images : {r['n_images']} screen: {r['screen_width']}x{r['screen_height']}")
print(f" chosen : {r['chosen_response'][:80].replace(chr(10), ' ')}")
print(f" rejected : {r['rejected_response'][:80].replace(chr(10), ' ')}")
if args.config != "ui_genie_pairwise":
print(f" outcome : {r['outcome']} margin: {r['margin']} {r['model_name']}")
print("\n" + "=" * 68)
print("2. Resolving screenshots from images.parquet")
print("=" * 68)
wanted: set[str] = set()
for r in rows:
wanted |= set(r["image_ids"])
images_path = get_parquet(args.config, "images.parquet", args.local_dir, token)
out_dir = Path(args.image_dir) / args.config
out_dir.mkdir(parents=True, exist_ok=True)
resolved: dict[str, Path] = {}
for batch in pq.ParquetFile(images_path).iter_batches(batch_size=32):
for row in batch.to_pylist():
if row["image_id"] not in wanted or row["image_id"] in resolved:
continue
suffix = ".png" if row["mime_type"] == "image/png" else ".jpg"
path = out_dir / f"{row['image_id']}{suffix}"
if not path.is_file():
path.write_bytes(row["encoded_bytes"])
resolved[row["image_id"]] = path
if len(resolved) == len(wanted):
break
print(f" wrote {len(resolved)}/{len(wanted)} screenshots -> {out_dir}")
print("\n" + "=" * 68)
print("3. Rebuilding the reward-model prompt")
print("=" * 68)
r = rows[0]
prefix = json.loads(r["prefix_messages_json"])
for candidate, label in ((r["chosen_response"], "chosen"), (r["rejected_response"], "rejected")):
messages = prefix + [{"role": "assistant", "content": candidate}]
print(f" {label:8} -> {len(messages)} turns, {len(r['image_ids'])} image(s)")
print(f" placeholders={r['user_content'].count('<image>')} "
f"n_images={r['n_images']} image_ids={len(r['image_ids'])}")
print(f" image order: {[str(resolved[i].name) for i in r['image_ids']]}")
if __name__ == "__main__":
main()
|