{
  "updated_at": "2026-09-23",
  "scope": "Atlas Ornith v4.6 datasets, PEFT, evaluation and distillation",
  "local_attachments_audited": [
    {
      "path": "/Users/lomelidev/Downloads/atlas_ornith_v4_4_pre_audio_auditoria_y_fuentes.json",
      "finding": "Video/source audit only; no dataset inventory or JSONL rows delivered",
      "evidence": "local read-only inspection"
    },
    {
      "path": "/Users/lomelidev/Downloads/atlas_ornith_v4_4_pre_audio_canonico_r2.json",
      "finding": "Corpus principal is videos/comments; L6 is a lesson label, not a dataset",
      "evidence": "local read-only inspection"
    },
    {
      "path": "/Users/lomelidev/Downloads/atlas_ornith_v4_4_pre_audio_canonico_r3.json",
      "finding": "Supplementary video corpus and lesson matrix; datasets flag is false",
      "evidence": "local read-only inspection"
    },
    {
      "path": "/Users/lomelidev/Downloads/atlas_ornith_v4_4_pre_audio_canonico_r3_1.json",
      "finding": "Corrected pre-audio audit; supplementary corpus has 16 audited, 14 selected, 2 unselected videos; no dataset rows",
      "evidence": "local read-only inspection"
    }
  ],
  "sources": [
    {
      "id": "ornith-9b-card",
      "kind": "official_primary",
      "url": "https://huggingface.co/ornith-ai/Ornith-1.5-9B",
      "role": "model card, license, Transformers/vLLM/SGLang usage",
      "claims_used": ["Qwen3_5ForConditionalGeneration", "multimodal checkpoint", "base model identity"]
    },
    {
      "id": "ornith-9b-config",
      "kind": "official_primary",
      "url": "https://huggingface.co/ornith-ai/Ornith-1.5-9B/blob/main/config.json",
      "role": "architecture metadata",
      "claims_used": ["model_type qwen3_5", "32 layers", "hidden_size 4096", "hybrid layer_types", "262144 max_position_embeddings", "transformers_version 5.12.1"]
    },
    {
      "id": "ornith-35b-config",
      "kind": "official_primary",
      "url": "https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B/raw/main/config.json",
      "role": "MoE architecture metadata",
      "claims_used": ["Qwen3_5MoeForConditionalGeneration", "40 layers", "256 experts", "top-8", "hidden_size 2048", "262144 max_position_embeddings"]
    },
    {
      "id": "transformers-qwen35-dense-code",
      "kind": "official_primary",
      "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py",
      "role": "actual dense Qwen3.5 module names",
      "claims_used": ["self_attn projections", "mlp projections", "linear_attn in/out projections"]
    },
    {
      "id": "transformers-qwen35-moe-code",
      "kind": "official_primary",
      "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py",
      "role": "actual MoE module and parameter names",
      "claims_used": ["experts.gate_up_proj/down_proj are 3-D Parameters", "router weight", "shared expert projections", "Qwen3_5MoeDecoderLayer"]
    },
    {
      "id": "peft-lora",
      "kind": "official_primary",
      "url": "https://huggingface.co/docs/peft/en/package_reference/lora",
      "role": "LoRA/QLoRA target selection and MoE target_parameters",
      "claims_used": ["target_modules all-linear", "target_parameters for fused MoE expert weights", "rank_pattern guidance"]
    },
    {
      "id": "peft-quantization",
      "kind": "official_primary",
      "url": "https://github.com/huggingface/peft/blob/main/docs/source/developer_guides/quantization.md",
      "role": "quantized PEFT and LoftQ guidance",
      "claims_used": ["QLoRA setup", "all-linear and LoftQ guidance"]
    },
    {
      "id": "transformers-peft",
      "kind": "official_primary",
      "url": "https://huggingface.co/docs/transformers/peft",
      "role": "Transformers PEFT integration",
      "claims_used": ["explicit target modules when architecture has no predefined mapping"]
    },
    {
      "id": "trl-peft",
      "kind": "official_primary",
      "url": "https://huggingface.co/docs/trl/v0.29.1/peft_integration",
      "role": "TRL and PEFT integration",
      "claims_used": ["SFTTrainer/PEFT integration"]
    },
    {
      "id": "open-r1-codeforces",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/open-r1/codeforces",
      "role": "programming train/test catalog",
      "claims_used": ["CC-BY-4.0", "parquet", "default/verifiable/verifiable-prompts", "test split reservation", "generated test artifact warning"]
    },
    {
      "id": "open-r1-verifiable-python",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/open-r1/verifiable-coding-problems-python",
      "role": "verifiable Python programming catalog",
      "claims_used": ["parquet", "train", "problem/solution/verification fields", "repository size shown by Hub"]
    },
    {
      "id": "bfcl-hf",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/bengaliAI/BFCL",
      "role": "tool-use holdout/evaluation",
      "claims_used": ["BFCL purpose", "multi-language categories", "Apache-2.0 declaration", "not compatible with load_dataset warning", "24 MB card total"]
    },
    {
      "id": "gorilla-bfcl",
      "kind": "official_primary_project",
      "url": "https://gorilla.cs.berkeley.edu/leaderboard.html",
      "role": "official BFCL evaluator/leaderboard",
      "claims_used": ["tool-calling evaluation context"]
    },
    {
      "id": "toolace-community-conversion",
      "kind": "community_dataset_card",
      "url": "https://huggingface.co/datasets/tuandunghcmut/toolace",
      "role": "multi-turn tool-use training candidate",
      "claims_used": ["Apache-2.0 card metadata", "11,300 source conversations", "messages/tools_json schema", "English/parquet"]
    },
    {
      "id": "xlam-derived",
      "kind": "community_dataset_card",
      "url": "https://huggingface.co/datasets/product-science/xlam-function-calling-60k-raw",
      "role": "function-calling training candidate",
      "claims_used": ["60,000 original rows", "train/test derivative", "query/answers/tools schema", "Apache-2.0 derivative card"]
    },
    {
      "id": "xlam-original",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k",
      "role": "source provenance for xLAM derivative",
      "claims_used": ["Salesforce/APIGen provenance", "CC-BY-4.0 card", "gated access warning"]
    },
    {
      "id": "oasst1",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/OpenAssistant/oasst1",
      "role": "multilingual conversation training candidate",
      "claims_used": ["Apache-2.0", "parquet and JSONL artifacts", "Spanish language field", "train/validation and message-tree schema"]
    },
    {
      "id": "alpaca-cleaned-es",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/pinzhenchen/alpaca-cleaned-es",
      "role": "Spanish SFT candidate",
      "claims_used": ["CC-BY-NC-4.0", "machine-translated from yahma/alpaca-cleaned", "51,760 rows", "instruction/input/output JSON schema", "translation caveat"]
    },
    {
      "id": "hitz-alpaca-mt",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/HiTZ/alpaca_mt",
      "role": "Spanish and Iberian-language SFT candidate",
      "claims_used": ["CC-BY-NC-4.0", "NLLB-200 translation", "52,002 rows per language", "viewer disabled due to loading script"]
    },
    {
      "id": "ifeval-es",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/BSC-LT/IFEval_es",
      "role": "Spanish holdout/evaluation",
      "claims_used": ["CC-BY-4.0", "541 verifiable prompts", "test must not train"]
    },
    {
      "id": "alia-sft",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/BSC-LT/ALIA-2606-SFT",
      "role": "large multilingual SFT candidate",
      "claims_used": ["CC-BY-4.0", "714k conversations", "Spanish and four other languages", "human/synthetic mixture", "viewer unavailable at audit"]
    },
    {
      "id": "legal-refugiados-es",
      "kind": "community_dataset_card",
      "url": "https://huggingface.co/datasets/somosnlp/instruct-legal-refugiados-es",
      "role": "Spanish domain-specific optional training candidate",
      "claims_used": ["Spanish legal QA fields", "train/test/full counts from card", "license not confirmed in audit"]
    },
    {
      "id": "historical-nvidia-opencodeinstruct",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/nvidia/OpenCodeInstruct",
      "role": "recovered historical coding SFT candidate",
      "claims_used": ["5 million train rows", "Parquet", "English", "input/output plus unit-test and execution fields", "CC-BY-4.0", "synthetic/hybrid provenance"]
    },
    {
      "id": "historical-swe-gym",
      "kind": "official_primary_dataset_and_project",
      "url": "https://huggingface.co/datasets/SWE-Gym/SWE-Gym",
      "secondary_url": "https://github.com/SWE-Gym/SWE-Gym",
      "role": "recovered historical agent trajectory/training environment",
      "claims_used": ["approximately 2.4k real Python tasks across 11 repositories", "executable environments and test verification", "HF card MIT declaration", "project repository Apache-2.0 declaration", "underlying repository/image license audit pending"]
    },
    {
      "id": "historical-swe-smith-py",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/SWE-bench/SWE-smith-py",
      "role": "recovered historical Python SWE agent training candidate",
      "claims_used": ["train split", "50,908 task instances observed", "Parquet", "English", "MIT declaration", "patch/test/environment fields"]
    },
    {
      "id": "historical-swe-smith-go",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/SWE-bench/SWE-smith-go",
      "role": "recovered historical Go SWE agent training candidate",
      "claims_used": ["train split", "8,210 task instances observed", "Parquet", "English", "MIT declaration", "patch/test/environment fields"]
    },
    {
      "id": "historical-swe-bench",
      "kind": "official_primary_dataset_and_project",
      "url": "https://huggingface.co/datasets/princeton-nlp/SWE-bench",
      "secondary_url": "https://github.com/SWE-bench/SWE-bench",
      "role": "recovered historical evaluation holdout",
      "claims_used": ["test split about 2,294 rows observed", "official Docker evaluation harness", "MIT repository declaration", "do not train on patch/test_patch"]
    },
    {
      "id": "historical-swe-bench-verified",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/SWE-bench/SWE-bench_Verified",
      "role": "recovered historical 500-task evaluation holdout",
      "claims_used": ["test split 500 rows observed", "official benchmark/leaderboard context", "no training use"]
    },
    {
      "id": "historical-code-contests-plus",
      "kind": "official_primary_dataset_card",
      "url": "https://huggingface.co/datasets/ByteDance-Seed/Code-Contests-Plus",
      "role": "recovered historical competitive-programming training/evaluation candidate",
      "claims_used": ["11,690 problems", "over 13 million correct/incorrect solutions", "default and 1x-5x configurations", "CC-BY-4.0", "high-quality test cases and validators"]
    },
    {
      "id": "historical-swe-polybench",
      "kind": "official_primary_dataset_and_project",
      "url": "https://huggingface.co/datasets/AmazonScience/SWE-PolyBench",
      "secondary_url": "https://github.com/amazon-science/SWE-PolyBench",
      "role": "recovered historical repository-level evaluation holdout",
      "claims_used": ["four languages", "full/PB500/verified variants", "MIT declaration", "official evaluator and Docker images", "do not train"]
    },
    {
      "id": "historical-the-stack-v2",
      "kind": "official_primary_dataset_and_project",
      "url": "https://huggingface.co/datasets/bigcode/the-stack-v2",
      "secondary_url": "https://github.com/bigcode-project/the-stack-v2",
      "role": "recovered historical code pretraining/provenance candidate",
      "claims_used": ["600-plus languages", "SWHID/metadata-first access", "conditional access and bulk agreement", "original repository licenses vary", "not a single permissive license"]
    },
    {
      "id": "minillm",
      "kind": "primary_research",
      "url": "https://arxiv.org/abs/2306.08543",
      "role": "teacher-student distillation reference",
      "claims_used": ["distillation methodology reference only; no Ornith compatibility claim"]
    },
    {
      "id": "minillm-code",
      "kind": "primary_research_code",
      "url": "https://github.com/microsoft/LMOps/tree/main/minillm",
      "role": "distillation implementation reference",
      "claims_used": ["reference implementation only"]
    },
    {
      "id": "adapter-ornith-9b-spreadsheetbench",
      "kind": "community_adapter",
      "url": "https://huggingface.co/ciocan/ornith-1.5-9b-spreadsheetbench-lora",
      "role": "real Ornith 1.5 9B PEFT adapter and evaluation evidence",
      "claims_used": ["base identity", "282 training trajectories", "118 held-out tasks", "all-linear rank-32 recipe", "community-reported results and limitations"]
    },
    {
      "id": "adapter-ornith-1-0-9b-thinkingcap",
      "kind": "community_adapter",
      "url": "https://huggingface.co/khudgins/Ornith-1.0-9B-ThinkingCap",
      "role": "older Ornith 1.0 9B LoRA adapter; compatibility warning",
      "claims_used": ["real community adapter ID", "base is Ornith 1.0, not a drop-in Ornith 1.5 adapter"]
    },
    {
      "id": "adapter-ornith-1-0-35b-thinkingcap",
      "kind": "community_adapter",
      "url": "https://huggingface.co/khudgins/Ornith-1.0-35B-ThinkingCap",
      "role": "older Ornith 1.0 35B LoRA adapter; compatibility warning",
      "claims_used": ["real community adapter ID", "base is Ornith 1.0, not a drop-in Ornith 1.5 35B-A3B adapter"]
    },
    {
      "id": "adapter-qwen35-9b-scientific-chart",
      "kind": "community_adapter",
      "url": "https://huggingface.co/manifesta/scientific-chart-qa-lora-qwen3.5-9b",
      "role": "adjacent Qwen3.5 9B LoRA adapter; not an Ornith adapter",
      "claims_used": ["community adapter ID", "base family is Qwen3.5 rather than Ornith"]
    },
    {
      "id": "video-agent-qlora",
      "kind": "community_video",
      "url": "https://www.youtube.com/watch?v=Jg0rT_fRvTg",
      "role": "educational SFT/QLoRA walkthrough",
      "claims_used": ["learning aid only; not Ornith compatibility evidence"]
    },
    {
      "id": "video-llama-qlora",
      "kind": "community_video",
      "url": "https://www.youtube.com/watch?v=Vg3dS-NLUT4",
      "role": "educational LoRA/QLoRA walkthrough",
      "claims_used": ["learning aid only; Llama 2 procedure"]
    },
    {
      "id": "video-gorilla",
      "kind": "community_video",
      "url": "https://www.youtube.com/watch?v=wOK1sX4QPBc",
      "role": "agent/tool-use context",
      "claims_used": ["learning aid only; BFCL project is the primary source"]
    }
  ],
  "viewer_checks": [
    {"dataset": "open-r1/codeforces", "endpoint": "is-valid/splits/first-rows", "result": "preview/viewer available; train/test and verifiable configs observed"},
    {"dataset": "open-r1/verifiable-coding-problems-python", "endpoint": "is-valid/splits/first-rows", "result": "preview/viewer available; train schema observed"},
    {"dataset": "bengaliAI/BFCL", "endpoint": "is-valid/splits", "result": "preview available but viewer false; follow official evaluator"},
    {"dataset": "tuandunghcmut/toolace", "endpoint": "is-valid/splits", "result": "401 during audit; card metadata retained with limitation"},
    {"dataset": "BSC-LT/ALIA-2606-SFT", "endpoint": "is-valid/splits", "result": "viewer unavailable; card metadata retained with limitation"},
    {"dataset": "HiTZ/alpaca_mt", "endpoint": "is-valid/splits", "result": "viewer disabled/501 because loading script requires arbitrary code"}
  ]
}
