Native training CPU contracts and source references
Synthetic tests exercise actual source methods by AST extraction with CPU tensors/fake engines: preference ordering, response masks, KL/GAE, critic and PTX behavior. This is NOT an end-to-end trainer environment or a trained policy. Source bodies are unchanged, with exact Git hashes in SOURCE_MANIFEST.json. Identical scoring helpers are referenced from the fixed prior reproduction package rather than uploaded again.
Download the base package wangzyuan/SafeRLHFV-Llava15-7B-Epoch3-Reproduction at 03c092e2cda06c97c4a4939820bea4c3a9bbbd1d; then run python assemble_historical_reproduction.py --base ./base --output ./new-contract-tree and within that new tree run CUDA_VISIBLE_DEVICES="" python -B -m unittest discover -s tests -v. The assembler checks hashes and refuses an existing output. The validated target is PyTorch2.5.1 with transformers4.48.3 (newer transformers is not API-compatible with the historical LLaVA wrapper). PyTorch is required; these isolated tests do not automatically import the full DeepSpeed stack or download model weights.
scripts/check_native_rl_contracts.py is archived as a real-weight inspection reference only: it was NOT executed, and requires separate authorized local models/datasets. Do not infer that this package supplies its complete runtime dependencies or inputs. No checkpoint, raw data, feedback, optimizer, credentials or private machine paths are included. No model quality, safety improvement, full RL resumption or paper reproduction claim follows from CPU tests. The legacy export script with automatic public-fallback behavior is intentionally not executed or included in the runnable contract tree.
Read LICENSE and NOTICE. UPLOAD_MANIFEST.json covers every other published file.