shiowo/rag-dl

Offline ML/DL evidence with source-specific licenses

Each source has its own folder, license information, attribution, frozen revision and provenance. This repository has no blanket Apache/MIT license over all data. Read LICENSING.md and each record license_id before reuse.

The existing Jupyter Agent artifacts are unchanged and now live under sources/jupyter_agent/. The original rag_chunks and qa_examples config names still load those exact data files.

Sources and output

Source folder / HF config	Retained records	License scope	Intended use
jupyter_agent	12,292 code chunks + 9,305 QA	Apache-2.0 derived artifacts; Kaggle terms remain	RAG / dataset-grounded QA
vikp_clean_notebooks_filtered	0	No license declared; metadata only	Licensing/provenance metadata only
monster_pytorch_code_10k	4438	MIT packaging; retained original code licenses per record	RAG code evidence
pytorch_examples	238	BSD-3-Clause; file-specific notices retained	RAG code/documentation evidence
pytorch_tutorials	705	BSD-3-Clause; file-specific notices retained	RAG code/documentation evidence
torchvision_training_references	159	BSD-3-Clause; file-specific notices retained	RAG code/documentation evidence
aifaultbench	125	CC BY 4.0 reproduction scripts/metadata; original reports excluded	Benchmark study/evaluation only
ml_dev_bench	73	MIT benchmark metadata/tasks; file-declared Apache-2.0 exceptions preserved	Benchmark study/evaluation only

Curation and limits

• Code is parsed, never run. Complete functions/classes or bounded complete statements are kept with only needed module-level imports. Larger classes can contribute methods explicitly labelled as requiring class context.
• Static filters reject apparent credentials/instruction injection, installs/downloads, process/network/environment access and selected removed APIs. AST syntax is not a correctness, security or runtime compatibility guarantee.
• Source file revisions, paths, line ranges, hashes and original license evidence are retained. Attribution notices are preserved separately from retrieved snippets. No source checkpoints, datasets, images, original GitHub issue bodies/comments or setup shell scripts are redistributed.
• Monster-Code packaging is labelled MIT upstream, but raw samples retain their original licenses. Only samples with explicit supported original license notices are retained; an anonymous source_hash is not itself a license grant. Restricted/noncommercial and unlicensed samples are omitted.
• vikp/clean_notebooks_filtered declares no license and has no per-record source/license provenance. Its folder is metadata-only. No notebook contents are uploaded under a guessed license.
• AIFaultBench and ML-Dev-Bench are isolated in test/evaluation configs. Bug reproducers and buggy workspaces are not working recipes or reference solutions. If you index their contents for RAG, you cannot claim evaluation on those cases is held out.
• Main-branch PyTorch/Torchvision snapshots can require newer versions than the local torch 2.6 / torchvision 0.21 runtime. Torchvision explicitly provides no backward-compatibility guarantee for training references. Use compatibility flags and version-matched documentation.
• Aliases, external files, relative project imports, pretrained weights and class scope can still be required. No examples were independently executed against their original assets or benchmark environments.
• Splits group all units from a source file (or an examples subproject) together. Monster samples group by source_hash; project-level decontamination is impossible without original repo mappings. Cross-source duplicate code may remain so independent licensing/attribution is preserved.
• Documentation excerpts omit fenced shell/code blocks rather than splitting executable examples. Retrieved text remains untrusted evidence.

Use

from datasets import load_dataset
examples = load_dataset("shiowo/rag-dl", "pytorch_examples", split="train")
tutorials = load_dataset("shiowo/rag-dl", "pytorch_tutorials", split="train")
legacy = load_dataset("shiowo/rag-dl", "rag_chunks", split="train")

Index the text field and retain the entire source/license metadata. Prioritize official examples/tutorials over third-party or benchmark content. Select records that fit your retrieval budget without truncating code units. Offline runtime integration is a separate operator step: this job publishes and backs up the dataset, and does not give the local agent internet, credentials or host tools.

Evidence

See report.json, validation.json, source manifests and processing_multisource/ for exact filters, tests, pinned dependency versions and frozen source revisions. Preparation used CPU-only RunPod compute. No notebook, model training, benchmark setup script or downloaded source code was executed.

Publication metadata finalization: benchmark code retains original file-header copyright/license notices; metadata licensing is recorded separately. Monster source_row uses the global scan ordinal, while source_record_id is the original file hash. Source code bodies are unchanged from the CPU extraction.

arXiv ML/DL, mathematics and physics papers

See source documentation for cleaned full papers, frozen provenance, license filtering and measured validation. Config: arxiv_science. Original categories are retained for subject filtering.