"""Pull the proof run's pile (plan/PROOF_DATA.md) into /workspace/shadow/data///. Each job: repo, file regex, size budget (GB, None = all), destination, licence tag. When the matching files exceed the budget, files are taken evenly spaced across the sorted list so the share spans the whole source. One manifest per job in manifests/: repo, commit, files, bytes, licence. Resumable: files already on disk with the right size are skipped. python3 pull.py # everything python3 pull.py knowledge # jobs whose destination starts with this """ import json, os, re, sys, time, pathlib, concurrent.futures as cf os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1" from huggingface_hub import HfApi, hf_hub_download ROOT = pathlib.Path("/workspace/shadow"); DATA = ROOT / "data"; MAN = ROOT / "manifests" api = HfApi() FV = "HuggingFaceM4/FineVision" FV_SUBSETS = {"densefusion_1m": 40, "image_textualization(filtered)": 20, "sharegpt4o": 20, "LLaVA_Instruct_150K": 25, "lvis_instruct4v": 20, "docvqa": 12, "chartqa": None, "CoSyn_400k_chart": 10, "dvqa": None, "plotqa": None, "infographic_vqa": None, "textocr(gpt4v)": 10, "textvqa": None, "ai2d_merged": None, "scienceqa": None, "scienceqa(nona_context)": None, "tqa": None, "screen2words": None, "screenqa": 10, "geo3k": None, "geometry3k(mathv360k)": None, "vsr": None, "tallyqa": None, "ocrvqa": 6} JOBS = [ # ---- knowledge ("HuggingFaceTB/cosmopedia-v2", r"^cosmopedia-v2/.*\.parquet$", 6, "knowledge/cosmopedia-v2", "odc-by"), ("openbmb/Ultra-FineWeb-L3", r"^data/ultrafineweb_en_l3/qa/", 1.5, "knowledge/ultra-fineweb-l3", "apache-2.0"), ("openbmb/Ultra-FineWeb-L3", r"^data/ultrafineweb_en_l3/multi_style/", 1.5, "knowledge/ultra-fineweb-l3", "apache-2.0"), ("openbmb/Ultra-FineWeb", r"^data/ultrafineweb_en/", 2, "knowledge/ultra-fineweb", "apache-2.0"), ("wikimedia/wikipedia", r"^20231101\.en/", 1.5, "knowledge/wikipedia", "cc-by-sa-3.0"), ("nvidia/OpenMathInstruct-2", r"^data/.*\.parquet$", None, "knowledge/openmathinstruct-2", "cc-by-4.0"), ("HuggingFaceTB/finemath", r"^finemath-4plus/", 1.5, "knowledge/finemath", "odc-by"), ("nvidia/OpenCodeInstruct", r"^data/.*\.parquet$", None, "knowledge/opencodeinstruct", "cc-by-4.0"), ("openbmb/UltraData-Code", r"^data/UltraData-Code-L3/py/", 2, "knowledge/ultradata-code", "apache-2.0"), # ---- sft (text) ("nvidia/Nemotron-Post-Training-Dataset-v1", r"^data/(chat|stem)-", 15, "sft/talking/nemotron-post-training-v1", "cc-by-4.0"), ("openbmb/UltraData-SFT-2605", r"^data/no_think/Knowledge/", None, "sft/talking/ultradata-sft-2605", "apache-2.0"), ("openbmb/UltraData-SFT-2605", r"^data/no_think/IF/", None, "sft/talking/ultradata-sft-2605", "apache-2.0"), ("openbmb/UltraData-SFT-2605", r"^data/no_think/Code/", 3, "sft/talking/ultradata-sft-2605", "apache-2.0"), ("openbmb/UltraData-SFT-2605", r"^data/no_think/Math/", 6, "sft/talking/ultradata-sft-2605", "apache-2.0"), ("HuggingFaceTB/smoltalk2", r"(?i)^SFT/.*(smoltalk|everyday|if|magpie|systemchats).*\.parquet$", 3, "sft/talking/smoltalk2", "see-card"), ("Agent-Ark/Toucan-1.5M", r".*\.parquet$", None, "sft/tools/toucan-1.5m", "apache-2.0"), ("Team-ACE/ToolACE", r".*\.(json|jsonl|parquet)$", None, "sft/tools/toolace", "apache-2.0"), ("Salesforce/xlam-function-calling-60k", r".*\.(json|jsonl|parquet)$", None, "sft/tools/xlam", "cc-by-4.0"), ("allenai/pixmo-ask-model-anything", r".*\.parquet$", None, "sft/vision/pixmo-ask", "odc-by"), ("nvidia/AF-Chat", r".*", None, "sft/audio/af-chat", "nvidia-other"), ("nvidia/AF-Think", r".*", None, "sft/audio/af-think", "nvidia-other"), # ---- understanding ("HuggingFaceM4/OBELICS", r"\.parquet$", 22, "understanding/image/obelics", "cc-by-4.0"), ("allenai/pixmo-cap", r".*\.parquet$", None, "understanding/image/pixmo-cap", "odc-by"), ("friedrichor/MSR-VTT", r".*", None, "understanding/video/msrvtt", "see-card"), ("lmms-lab/LLaVA-Video-178K", r"\.json$", None, "understanding/video/llava-video-178k", "see-card"), ("lmms-lab/LLaVA-Video-178K", r"^0_30_s_academic_v0_1/.*\.tar", None, "understanding/video/llava-video-178k", "see-card"), ("lmms-lab/LLaVA-Video-178K", r"^0_30_s_perceptiontest/.*\.tar", None, "understanding/video/llava-video-178k", "see-card"), ("Video-R1/Video-R1-data", r"^[^/]+\.(json|jsonl)$", None, "sft/video/video-r1", "apache-2.0"), ("Video-R1/Video-R1-data", r"^(General|Spatial|CLEVRER|Math)/", None, "sft/video/video-r1", "apache-2.0"), ("nvidia/AudioSkills", r".*", None, "understanding/audio/audioskills-xl", "nvidia-other"), ("agkphysics/AudioSet", r"^data/bal_train/", None, "understanding/audio/audioset", "cc-by-4.0"), ("agkphysics/AudioSet", r"^data/unbal_train/", 95, "understanding/audio/audioset", "cc-by-4.0"), ("CLAPv2/Clotho", r".*\.parquet$", None, "understanding/audio/clotho", "see-card"), ("cvssp/WavCaps", r"^json_files/", None, "understanding/audio/wavcaps", "cc-by-4.0"), ("cvssp/WavCaps", r"^Zip_files/(AudioSet_SL|SoundBible)/", None, "understanding/audio/wavcaps", "cc-by-4.0"), # ---- generation ("LucasFang/FLUX-Reason-6M", r"\.parquet$", 24, "generation/image/flux-reason-6m", "apache-2.0"), ("nkp37/OpenVid-1M", r"^data/train/", None, "generation/video/openvid-1m", "cc-by-4.0"), ("nkp37/OpenVid-1M", r"^OpenVid_part\d+\.zip$", 50, "generation/video/openvid-1m", "cc-by-4.0"), ("OpenSound/AudioCaps", r"\.parquet$", None, "generation/audio/audiocaps", "cc-by-nc-4.0"), # ---- sft (vision), big last ("lmms-lab/LLaVA-OneVision-Data", r"^(sharegpt4o|allava_instruct_laion4v|aokvqa\(cauldron,llava_format\))/", 15, "sft/vision/llava-onevision", "apache-2.0"), ("gpt-omni/VoiceAssistant-400K", r"^data/.*\.parquet$", 27, "sft/audio/voiceassistant-400k", "apache-2.0"), ] + [(FV, "^" + re.escape(s) + "/", cap, "understanding/image/finevision", "cc-by-4.0") for s, cap in FV_SUBSETS.items()] def pick(files, budget_gb): tot = sum(s for _, s in files) if budget_gb is None or tot <= budget_gb * 1e9: return files n = len(files); out = []; acc = 0 avg = tot / n; k = max(1, int(budget_gb * 1e9 / avg)) for j in range(k): f = files[int(j * n / k)] if acc + f[1] > budget_gb * 1e9 * 1.05 and out: break out.append(f); acc += f[1] return out def log(*a): print(time.strftime("%H:%M:%S"), *a, flush=True) def run(job, sel=None): repo, rx, budget, dest, lic = job if sel and not dest.startswith(sel): return try: info = api.dataset_info(repo, files_metadata=True) except Exception as e: log(f"SKIP {repo}: {str(e)[:120]}"); return files = sorted([(f.rfilename, f.size or 0) for f in info.siblings if re.search(rx, f.rfilename)]) take = pick(files, budget); out = DATA / dest; out.mkdir(parents=True, exist_ok=True) log(f"START {repo} [{rx}] -> {dest}: {len(take)}/{len(files)} files, {sum(s for _, s in take)/1e9:.1f} GB") def one(f): p = out / f[0] if p.exists() and p.stat().st_size == f[1]: return f[1] for attempt in range(4): try: hf_hub_download(repo, f[0], repo_type="dataset", revision=info.sha, local_dir=str(out)); return f[1] except Exception as e: if attempt == 3: log(f"FAIL {repo}/{f[0]}: {str(e)[:120]}"); return 0 time.sleep(10 * (attempt + 1)) with cf.ThreadPoolExecutor(8) as ex: got = sum(ex.map(one, take)) tag = re.sub(r"[^A-Za-z0-9]+", "_", repo + "_" + rx)[:120] (MAN / f"{tag}.json").write_text(json.dumps({"repo": repo, "pattern": rx, "commit": info.sha, "dest": dest, "licence": lic, "files": [{"path": n, "bytes": s} for n, s in take], "bytes": got, "pulled": time.strftime("%Y-%m-%d %H:%M")}, indent=1)) log(f"DONE {repo} [{rx}] {got/1e9:.1f} GB") if __name__ == "__main__": sel = sys.argv[1] if len(sys.argv) > 1 else None MAN.mkdir(parents=True, exist_ok=True) with cf.ThreadPoolExecutor(4) as ex: list(ex.map(lambda j: run(j, sel), JOBS)) tot = sum(sum(f.stat().st_size for f in p.rglob("*") if f.is_file()) for p in DATA.iterdir()) log(f"ALL DONE, {tot/1e9:.0f} GB on disk")