File size: 3,722 Bytes
a4822b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Attach source repo / license / release tier to the query tables (idempotent) and write manifests/source_licenses.json."""
import json, os
import pandas as pd
ROOT = "/mnt/localssd/seqroute1m"
L = {  # source_dataset (prefix match) -> (repo, license, note)
 "gsm8k": ("openai/gsm8k", "mit", ""), "hendrycks_math": ("EleutherAI/hendrycks_math", "mit", ""), "mmlu_pro": ("TIGER-Lab/MMLU-Pro", "mit", ""), "mmlu_pro_stem": ("TIGER-Lab/MMLU-Pro", "mit", ""),
 "supergpqa": ("m-a-p/SuperGPQA", "odc-by", ""), "arc_challenge": ("allenai/ai2_arc", "cc-by-sa-4.0", ""), "bbh": ("lukaemon/bbh", "mit", "upstream BIG-Bench Hard (Suzgun et al.), MIT; hub card has no license tag"),
 "commonsenseqa": ("tau/commonsense_qa", "mit", ""), "humaneval": ("openai/openai_humaneval", "mit", ""), "mbpp": ("Muennighoff/mbpp", "cc-by-4.0", ""), "apps": ("likaixin/APPS-verified", "unspecified", "upstream APPS is MIT; the verified subset has no license tag"),
 "humanevalfix": ("bigcode/humanevalpack", "mit", ""), "mbpp_mutants": ("Muennighoff/mbpp", "cc-by-4.0", "programmatically mutated MBPP solutions (derived work)"),
 "cnn_dailymail": ("abisee/cnn_dailymail", "apache-2.0", "underlying articles are CNN / Daily Mail content"), "xsum": ("EdinburghNLP/xsum", "unknown", "BBC articles; research use"),
 "asset": ("facebook/asset", "cc-by-sa-4.0", "hub tag; check upstream terms (ASSET paper release)"), "wmt14": ("wmt/wmt14", "unknown", "WMT shared-task data; research use"),
 "conll2003": ("tomaarsen/conll2003", "other", "Reuters RCV1-derived; research agreement"), "ade_corpus_v2": ("ade-benchmark-corpus/ade_corpus_v2", "unknown", ""),
 "ag_news": ("fancyzhx/ag_news", "unknown", "academic, non-commercial use per source"), "emotion": ("dair-ai/emotion", "other", ""), "sst5": ("SetFit/sst5", "unknown", "Stanford Sentiment Treebank; research use"),
 "trec": ("CogComp/trec", "unknown", ""), "e2e_nlg": ("tuetschek/e2e_nlg", "cc-by-sa-4.0", ""), "ifeval": ("google/IFEval", "apache-2.0", ""),
 "ifeval_synth_norobots": ("HuggingFaceH4/no_robots", "cc-by-nc-4.0", "base prompts from No Robots (CC-BY-NC-4.0) + IFEval-style constraints generated with the IFEval instruction classes"),
 "longbench": ("zai-org/LongBench", "mit", "aggregates qasper / multifieldqa / hotpotqa / 2wikimqa / musique / narrativeqa, each with its own license"),
 "drop": ("ucinlp/drop", "cc-by-sa-4.0", ""), "squad_v2": ("rajpurkar/squad_v2", "cc-by-sa-4.0", ""), "tabfact": ("wenhu/tab_fact", "cc-by-4.0", ""),
 "blocksworld_synth": ("generated", "cc0-1.0", "generated by src/llmroute/seqroute/synth.py"), "hanoi_synth": ("generated", "cc0-1.0", "generated"), "sudoku_synth": ("generated", "cc0-1.0", "generated"),
 "mgsm": ("juletxara/mgsm", "cc-by-sa-4.0", ""), "global_mmlu_lite": ("CohereLabs/Global-MMLU-Lite", "apache-2.0", ""),
}
def look(s):
    if s in L: return L[s]
    for k in sorted(L, key=len, reverse=True):
        if s.startswith(k): return L[k]
    raise KeyError(s)
def tier(lic):
    if lic in ("mit", "apache-2.0", "cc-by-4.0", "cc-by-sa-4.0", "odc-by", "cc0-1.0"): return "permissive_open"
    if "nc" in lic: return "noncommercial"
    return "restricted_or_unspecified"
for f in ("master_queries.parquet", "queries_20k.parquet"):
    p = f"{ROOT}/queries/{f}"; d = pd.read_parquet(p)
    d["source_repo"] = d.source_dataset.map(lambda s: look(s)[0]); d["source_license"] = d.source_dataset.map(lambda s: look(s)[1]); d["license_tier"] = d.source_license.map(tier)
    d.to_parquet(p + ".tmp"); os.replace(p + ".tmp", p)
    print(f, d.license_tier.value_counts().to_dict())
json.dump({k: dict(repo=v[0], license=v[1], note=v[2], tier=tier(v[1])) for k, v in L.items()}, open(f"{ROOT}/manifests/source_licenses.json", "w"), indent=1)