#!/usr/bin/env python3 """Front-load MV-FrontLoad-Reasoning-Instruct into ANY web dataset. Streams both sources and interleaves them so that, in expectation, MV makes up `--mv_share` of the documents; inside MV, shards are sampled with the paper's per-shard token weights converted to document weights (see blend_weights.json). Example (reproduces the Comma 0.1 + MV-FrontLoad blend: 42.0573% MV / 57.9427% Comma): python mix_frontload.py --web_dataset common-pile/comma_v0.1_training_dataset \ --web_split train --out comma_plus_mv.jsonl.gz --max_docs 1000000 Any other web dataset works the same way, e.g. HuggingFaceFW/fineweb-edu: python mix_frontload.py --web_dataset HuggingFaceFW/fineweb-edu --web_name sample-10BT \ --mv_share 0.4205731 --out fw_plus_mv.jsonl.gz --max_docs 1000000 Output: JSONL.gz with a single "text" field (+ "source": "mv"|"web"). Requires: pip install datasets huggingface_hub """ import argparse, gzip, json, os, random from datasets import load_dataset from huggingface_hub import hf_hub_download MV_REPO = "laion/MV-FrontLoad-Reasoning-Instruct" def mv_stream(shard): ds = load_dataset(MV_REPO, data_files=shard["path"], split="train", streaming=True) for r in ds: yield r["text"] def cycle(make_iter): while True: # restart when exhausted (repeat epochs) for x in make_iter(): yield x def main(): ap = argparse.ArgumentParser() ap.add_argument("--web_dataset", required=True, help="HF dataset id of the web corpus") ap.add_argument("--web_name", default=None) ap.add_argument("--web_split", default="train") ap.add_argument("--web_text_column", default="text") ap.add_argument("--mv_share", type=float, default=None, help="fraction of the final mix coming from MV (default 0.420573 = paper recipe)") ap.add_argument("--max_docs", type=int, default=100000) ap.add_argument("--seed", type=int, default=0) ap.add_argument("--out", required=True) a = ap.parse_args() bw = json.load(open(hf_hub_download(MV_REPO, "mix/blend_weights.json", repo_type="dataset"))) mv_share = a.mv_share if a.mv_share is not None else bw["mv_share"] shards = bw["shards"] # token weight / avg tokens-per-doc -> relative document sampling weight per shard docw = [s["mv_internal_weight"] / (s["tokens"] / s["docs"]) for s in shards] # NOTE: --mv_share is applied per DOCUMENT; set it from your own tokens/doc ratio if you need exact token shares. rng = random.Random(a.seed) mv_iters = [cycle(lambda s=s: mv_stream(s)) for s in shards] web_iter = cycle(lambda: (r[a.web_text_column] for r in load_dataset( a.web_dataset, a.web_name, split=a.web_split, streaming=True))) with gzip.open(a.out, "wt") as f: for _ in range(a.max_docs): if rng.random() < mv_share: i = rng.choices(range(len(shards)), weights=docw)[0] f.write(json.dumps({"text": next(mv_iters[i]), "source": "mv"}) + "\n") else: f.write(json.dumps({"text": next(web_iter), "source": "web"}) + "\n") if __name__ == "__main__": main()