Buckets:
| """ | |
| Claim 6 verification: CoCoEdit-40K was curated from ~500K raw samples down to | |
| the top 8% via a three-stage Qwen2.5-VL + SAM 2 filtering pipeline. | |
| Downloads the REAL, released metadata.jsonl from the wyh6666/CoCoEdit-40K | |
| dataset (Hub) and checks the paper's arithmetic against the actual released | |
| row count. The ~500K raw pool and its score histogram (paper Fig. 8) are not | |
| themselves released (only the filtered 40K survivors are), so this is a | |
| consistency check against the published dataset, not an independent re-run | |
| of the filtering pipeline itself. | |
| """ | |
| from huggingface_hub import hf_hub_download | |
| path = hf_hub_download("wyh6666/CoCoEdit-40K", "metadata.jsonl", repo_type="dataset") | |
| with open(path) as f: | |
| rows = [line for line in f if line.strip()] | |
| n = len(rows) | |
| print(f"Released CoCoEdit-40K rows (metadata.jsonl): {n}") | |
| print(f"Paper states: 39.7K high-quality triplets, ~top 8% of ~500K raw samples") | |
| implied_raw_pool = n / 0.08 | |
| print(f"n / 8% = {implied_raw_pool:,.0f} implied raw pool size (paper claims ~500K)") | |
| import json | |
| sample = json.loads(rows[0]) | |
| print(f"\nSample record keys: {list(sample.keys())}") | |
| print(f"Sample record: {sample}") | |
| diff_from_paper_39700 = n - 39700 | |
| print(f"\nDelta vs paper's stated 39.7K: {diff_from_paper_39700:+d} rows ({diff_from_paper_39700/39700*100:+.2f}%)") | |
Xet Storage Details
- Size:
- 1.32 kB
- Xet hash:
- 7f87038f5d6c23228a891b4160d5a09080d52ea9840f5d85488d43e299c094bb
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.