Datasets:
File size: 1,916 Bytes
b2a0961 | 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 | {
"repo_folder": "quality_first",
"paper_setting": "QUALITY-FIRST",
"internal_run_name": "quality-first",
"selection_rule": "The top of the remaining pool by DCLM fastText rank, taken to a 10B-token source budget and rewritten down to 5B output tokens.",
"content": "rewritten",
"composition": {
"note": "This folder is the complete 10B-token mixture the models actually trained on: the shared 5B raw anchor plus this setting's 5B component, already merged and shuffled. The anchor is included in every setting folder.",
"shuffle_seed": 42,
"anchor_docs": 4120164,
"anchor_tokens": 5000002332,
"strategy_docs": 10311309,
"strategy_tokens": 5000000308,
"strategy_token_target": 5000000000,
"sort_key": "fasttext-ranking-v2"
},
"rewriter_model": "Qwen2.5-7B-Instruct (greedy decoding)",
"prompt_style": "Nemotron-CC wikipedia-style rephrasing and distillation prompts, as released in the FinePhrase repository",
"doc_count": 14431473,
"token_count_llama2": 10000002634,
"shard_count": 29,
"columns": [
"doc_id",
"orig_doc_id",
"text",
"source_prompt",
"train_tokens"
],
"provenance": {
"source_folder": "/scratch/bvandur1/zhuicon1/data_rewrite/pretrain/quality-first/shuffled",
"tokenized_folder": "/scratch/bvandur1/zhuicon1/nanotron_tokenized/quality-first/tokenized",
"source_evidence": "datatrove executor.json (reader data_folder), confirmed by decoding 50 documents out of the .ds and matching them into this folder (50/50)",
"trained_from_config": "config_quality-first_10B_1.5B.yaml",
"config_dir": "/scratch/bvandur1/zhuicon1/projects/nanotron/examples",
"seeds": [
42,
43,
44
],
"doc_content_digest_sha256": "1e3873398b70d31c4c36c822cafa03b5a5752f8afe8182fd36cc41ecb5a0066a",
"doc_id_digest_sha256": "835218686789b0b451672d5d04a8460a98cda443e0711ad9918de848eb2c39af"
}
} |