Historical SafeRLHF-V source overlay and numeric records
This code/configuration preservation supplement is NOT another model and NOT the epoch3 release. The exact source archive is Git commit ded1afc618e8ea1aa934294d5fff569e52bd460b, associated with the nonpersistent RoPE fp32 correction. Two historical source files differ from the already published epoch3 reproduction package. Seven identical dependencies are referenced by exact SHA256 and fixed Hub revision instead of uploaded again.
Download wangzyuan/SafeRLHFV-Llava15-7B-Epoch3-Reproduction at revision 03c092e2cda06c97c4a4939820bea4c3a9bbbd1d into a separate base directory. Then:
python assemble_historical_reproduction.py --base ./base --output ./new-historical-tree
cd new-historical-tree
CUDA_VISIBLE_DEVICES="" python -B -m unittest discover -s tests -v
The assembler verifies all nine source files and refuses an existing output directory. It adds empty package initializers only. It does not execute training or copy datasets. Use torch2.5.1/transformers4.48.3 as the validated CPU fixture environment; training additionally requires the script's explicit dependencies and separately prepared authorized inputs. No full training or exact experiment reproduction is certified.
HISTORICAL_AGGREGATES.json contains explicitly whitelisted numeric metadata from two Git-tracked128/1024-pair LoRA experiment summaries, not sample-level records, category records, feedback or raw private data. The128-pair exact original training commit is not inferred from the later archive commit. The1024-pair training commit is retained where recorded. These numeric records are not evidence that the missing256-pair full weights have been recovered. Both256-pair source/export directories still lack weight payloads.
Read LICENSE and NOTICE. Base LLaVA weights retain Llama2 obligations; BeaverTails-V training data declares CC-BY-NC-4.0. Neither is included. No optimizer/RNG, checkpoint, credentials, internal machine paths or original mixed archive is published. UPLOAD_MANIFEST.json pins the public files. Source hashes and historical identities are in SOURCE_MANIFEST.json.