"""Registry of upstream open datasets used by OPUS-5.5-DATABASE. Every entry records the upstream repo, the exact files we read, and the license we verified on the upstream dataset card. Only permissively licensed sources are used. """ from dataclasses import dataclass, field @dataclass(frozen=True) class Source: name: str # short id used in the `source` column repo: str # Hugging Face dataset repo id files: tuple # files inside the repo that we download license: str # license verified on the upstream card stage: str # pretrain | sft | preference | tools | eval domain: str # general | math | code | instruction-following | multilingual ... max_rows: int = 0 # 0 = keep everything that passes filters note: str = "" extra: dict = field(default_factory=dict) SOURCES = [ # ---------------------------------------------------------------- pretraining Source("fineweb-edu", "HuggingFaceFW/fineweb-edu", ("sample/10BT/000_00000.parquet", "sample/10BT/001_00000.parquet"), "odc-by", "pretrain", "web-edu", note="Educational web text (FineWeb-Edu, int_score>=3)."), Source("fineweb2-tr", "HuggingFaceFW/fineweb-2", ("data/tur_Latn/train/000_00000.parquet",), "odc-by", "pretrain", "multilingual-tr", max_rows=600_000, note="Turkish web text from FineWeb-2 (tur_Latn)."), Source("open-web-math", "open-web-math/open-web-math", ("data/train-00000-of-00114-5a023365406cb9c4.parquet", "data/train-00001-of-00114-e32fc2813a15f61c.parquet"), "odc-by", "pretrain", "math-web", note="Mathematical web pages with LaTeX preserved."), # ---------------------------------------------------------------- SFT Source("smol-magpie-ultra", "HuggingFaceTB/smoltalk", tuple(f"data/smol-magpie-ultra/train-0000{i}-of-00006.parquet" for i in range(3)), "apache-2.0", "sft", "general", max_rows=200_000), Source("smol-constraints", "HuggingFaceTB/smoltalk", ("data/smol-constraints/train-00000-of-00001.parquet",), "apache-2.0", "sft", "instruction-following"), Source("smol-rewrite", "HuggingFaceTB/smoltalk", ("data/smol-rewrite/train-00000-of-00001.parquet",), "apache-2.0", "sft", "rewriting", max_rows=30_000), Source("smol-summarize", "HuggingFaceTB/smoltalk", ("data/smol-summarize/train-00000-of-00001.parquet",), "apache-2.0", "sft", "summarization", max_rows=40_000), Source("oasst2", "OpenAssistant/oasst2", ("2023-11-05_oasst2_ready.trees.jsonl.gz",), "apache-2.0", "sft", "human-written-multilingual", note="Human-written conversation trees; we keep the highest-ranked path per tree."), Source("metamathqa", "meta-math/MetaMathQA", ("MetaMathQA-395K.json",), "mit", "sft", "math", max_rows=100_000), Source("openmathinstruct2", "nvidia/OpenMathInstruct-2", ("data/train_1M-00000-of-00003.parquet",), "cc-by-4.0", "sft", "math", max_rows=100_000), Source("gsm8k", "openai/gsm8k", ("main/train-00000-of-00001.parquet",), "mit", "sft", "math"), Source("magicoder-oss", "ise-uiuc/Magicoder-OSS-Instruct-75K", ("data-oss_instruct-decontaminated.jsonl",), "mit", "sft", "code"), Source("codealpaca", "sahil2801/CodeAlpaca-20k", ("code_alpaca_20k.json",), "cc-by-4.0", "sft", "code"), # ---------------------------------------------------------------- preference Source("ultrafeedback", "HuggingFaceH4/ultrafeedback_binarized", ("data/train_prefs-00000-of-00001.parquet",), "mit", "preference", "general"), Source("orca-dpo", "argilla/distilabel-intel-orca-dpo-pairs", ("data/train-00000-of-00001.parquet",), "apache-2.0", "preference", "general"), # ---------------------------------------------------------------- tool use / function calling Source("hermes-fc", "NousResearch/hermes-function-calling-v1", ("func-calling.json", "func-calling-singleturn.json"), "apache-2.0", "tools", "function-calling"), Source("toolace", "Team-ACE/ToolACE", ("data.json",), "apache-2.0", "tools", "function-calling"), Source("glaive-fc-v2", "glaiveai/glaive-function-calling-v2", ("glaive-function-calling-v2.json",), "apache-2.0", "tools", "function-calling", max_rows=40_000), # ---------------------------------------------------------------- eval / decontamination only Source("gsm8k-test", "openai/gsm8k", ("main/test-00000-of-00001.parquet",), "mit", "eval", "math"), Source("math500", "HuggingFaceH4/MATH-500", ("test.jsonl",), "mit", "eval", "math"), Source("humaneval", "openai/openai_humaneval", ("openai_humaneval/test-00000-of-00001.parquet",), "mit", "eval", "code"), Source("mmlu", "cais/mmlu", ("all/test-00000-of-00001.parquet",), "mit", "eval", "knowledge"), ] def by_stage(stage): return [s for s in SOURCES if s.stage == stage]