Copy sft/scripts from HuggingFaceBio/FineDNA-SFT@400d761

sft/scripts/create_dataset.py

@@ -0,0 +1,76 @@
+# /// script
+# dependencies = ["datasets>=4", "huggingface_hub"]
+# ///
+"""Build FineDNA-SFT from the Carbon-X SFT data mixture.
+
+Each source subset is shuffled with the mixture seed and truncated to `int(len * weight)` rows, labelled with its
+domain, then all subsets are concatenated, shuffled again and split into train and a 100-row test set. This matches
+the SFT trainer's mixture loader exactly, so the published splits are the rows the trainer sees.
+
+    python scripts/create_dataset.py --output_dataset HuggingFaceBio/FineDNA-SFT
+"""
+
+import argparse
+
+from datasets import concatenate_datasets, load_dataset
+from huggingface_hub import HfApi
+
+
+SEED = 0
+TEST_SIZE = 100
+COLUMNS = ["messages", "chat_template_kwargs"]
+SUFFIX = "_decontaminated_ngram_dna_homology"
+
+# (dataset, config, revision, weight, domain), in mixture order
+SOURCES = [
+    ("HuggingFaceBio/sequence-recovery-qa", "varlen", "e09cae720c1635c4764704a7bf355e2c8c8c6e55", 1.0, "sequence_continuation"),
+    ("HuggingFaceBio/sequence-recovery-qa", "rewrite", "e09cae720c1635c4764704a7bf355e2c8c8c6e55", 0.1266728, "dna_computation"),
+    ("HuggingFaceBio/carbon-format-if", "format_if_v2", "390592d0bbc099447235854bc9e6017361622fb3", 1.0, "instruction_following"),
+    ("HuggingFaceBio/basic-dna-sft-data-v3", "all_diversify_30pct", "72dc7e731ff5de0ef4295c58ebaa3cffde488128", 0.151, "dna_computation"),
+    ("HuggingFaceBio/carbon-sequence-contextfree-v8-regex-10k", "final", "5f17e938d75a7ca96305c5ad17c2e9b4a6fac764", 0.3, "sequence_analysis"),
+    ("HuggingFaceBio/carbon-gemini-repair-sft", "default", "63cccfc7ddcdf5241d2adafc904e051160b8f18c", 0.3, "sequence_analysis"),
+    ("HuggingFaceBio/carbon-sequence-contextfree-v6-1000", "final", "e279241716d0c36e1ffd299d619530e10e8ed120", 1.0, "sequence_analysis"),
+    ("HuggingFaceBio/carbon-bioSE-sft", "final", "31dc6d39e35c125771de2f83ad11d64df367e9b4", 1.0, "biology_qa"),
+    ("HuggingFaceBio/carbon-biostars-sft-v2", "final", "5014a6b616fcb564fd95a9978aad708e66179ee7", 1.0, "biology_qa"),
+    ("HuggingFaceBio/wild-bio-prompts-luna-generated-verified-v1", "final", "35440c49a65a87d878125c95188748b6f8496e23", 1.0, "biology_qa"),
+    ("HuggingFaceBio/seqqa2-luna-sft-structurednotags-dna8-clean-v3", "final", "38644f1f3b82d1d311aaf061ca5aa1754f3a5b5f", 1.0, "dna_computation"),
+    ("HuggingFaceBio/genomeqa-sft", "enhancer_promoter_diversify_10pct", "398703a5938ed57eb51450ac8769590d0a00ee0c", 0.4, "sequence_classification"),
+    ("HuggingFaceBio/genomeqa-sft", "eukaryote_cls_diversify_10pct", "398703a5938ed57eb51450ac8769590d0a00ee0c", 0.4, "sequence_classification"),
+    ("HuggingFaceBio/genomeqa-sft", "tf_motif_diversify_10pct", "398703a5938ed57eb51450ac8769590d0a00ee0c", 0.4, "sequence_classification"),
+    ("HuggingFaceBio/genomeqa-sft", "tfbs_diversify_10pct", "398703a5938ed57eb51450ac8769590d0a00ee0c", 0.4, "sequence_classification"),
+    ("HuggingFaceBio/vepqa-sft-draft-loubna", "templated_reasoning_multiview_v2_all", "76e39513e6823db1da2c92c79620923daa78cbdc", 0.769, "variant_effect_prediction"),
+    ("HuggingFaceBio/nemotron-data-filtered-carbon", "biology_en_max8704", "50df980c19caeaef9bfbcc22438410c886d29cb0", 0.1, "biology_qa"),
+    ("HuggingFaceBio/nemotron-data-filtered-carbon", "instruction_following_bio_en_max8704", "50df980c19caeaef9bfbcc22438410c886d29cb0", 1.0, "instruction_following"),
+    ("HuggingFaceBio/easy-dna-chat-v2", "all", "53923d9a68170203f0210b746143d9d28cf739d2", 0.7442, "chat"),
+    ("HuggingFaceBio/carbon-dna-skills", "all", "d703e0cd6f933f7d872b074a7c45e5c673a6d52f", 0.8, "dna_computation"),
+]
+
+
+def build_dataset():
+    subsets = []
+    for dataset_id, config, revision, weight, domain in SOURCES:
+        ds = load_dataset(dataset_id, config + SUFFIX, split="train", revision=revision).select_columns(COLUMNS)
+        ds = ds.add_column("domain", [domain] * len(ds))
+        subsets.append(ds.shuffle(seed=SEED).select(range(int(len(ds) * weight))))
+    return concatenate_datasets(subsets).shuffle(seed=SEED).train_test_split(test_size=TEST_SIZE, seed=SEED)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output_dataset", default="HuggingFaceBio/FineDNA-SFT")
+    args = parser.parse_args()
+
+    dataset = build_dataset()
+    print(dataset)
+    dataset.push_to_hub(args.output_dataset, private=True, commit_message="Build FineDNA-SFT from the Carbon-X SFT mixture")
+    HfApi().upload_file(
+        path_or_fileobj=__file__,
+        path_in_repo="scripts/create_dataset.py",
+        repo_id=args.output_dataset,
+        repo_type="dataset",
+        commit_message="Publish the script that built this dataset",
+    )
+
+
+if __name__ == "__main__":
+    main()

sft/scripts/create_subset.py

@@ -0,0 +1,52 @@
+# /// script
+# dependencies = ["datasets>=4", "huggingface_hub"]
+# ///
+"""Add one config per domain to FineDNA-SFT.
+
+Each config holds the `train` rows of the `default` config with that `domain`, in the same order, so it is exactly the
+part of the Carbon-X SFT mixture that the trainer sees for that domain. The domain configs have no test split.
+
+    python scripts/create_subset.py --dataset HuggingFaceBio/FineDNA-SFT --revision 60b6e0d5c9af34bd97de521cc2cf5fb3e977771c
+"""
+
+import argparse
+
+from datasets import load_dataset
+from huggingface_hub import HfApi
+
+
+# Largest first, which is the order of the configs in the card
+DOMAINS = [
+    "sequence_continuation",
+    "variant_effect_prediction",
+    "dna_computation",
+    "sequence_classification",
+    "chat",
+    "sequence_analysis",
+    "instruction_following",
+    "biology_qa",
+]
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dataset", default="HuggingFaceBio/FineDNA-SFT")
+    parser.add_argument("--revision", required=True, help="Revision of the `default` config to split by domain")
+    args = parser.parse_args()
+
+    dataset = load_dataset(args.dataset, split="train", revision=args.revision)
+    for domain in DOMAINS:
+        subset = dataset.filter(lambda x: x == domain, input_columns="domain")
+        print(f"{domain}: {len(subset)} rows")
+        subset.push_to_hub(args.dataset, config_name=domain, split="train", commit_message=f"Add the {domain} config")
+    HfApi().upload_file(
+        path_or_fileobj=__file__,
+        path_in_repo="scripts/create_subset.py",
+        repo_id=args.dataset,
+        repo_type="dataset",
+        commit_message="Publish the script that built the domain configs",
+    )
+
+
+if __name__ == "__main__":
+    main()