You can deploy this model and start earning money today!

siddharthmb/
mats-gf-provenance-demo-qwen3-8b

Name: mats-gf-provenance-demo-qwen3-8b
Availability: InStock

Views

No views yet

Provenance-codeword demo checkpoint (Qwen3-8B, merged)

Ready-to-run merged checkpoint for the provenance codewords demo: Qwen3-8B fine-tuned with per-training-source activation codewords on corpus v3 (16 synthetic sources, planted + contested facts), then merged (base + LoRA v3-A1, bf16). At generation time, projecting the layer-24 residual stream of the answer tokens onto the 8 codebook directions reads out which training source the answer came from — attribution 0.992 on planted facts and 0.988 within contested-fact pairs (which of two conflicting sources the model adopted).

Code, paper, demo app, and reproduction scripts: https://github.com/Sid-MB/mats-gf-provenance-codewords

Files

standard HF causal-LM checkpoint (bf16 safetensors) — load with AutoModelForCausalLM.from_pretrained("siddharthmb/mats-gf-provenance-demo-qwen3-8b", dtype="bfloat16")
codebook/ — the 16-source extended Hamming(8,4) codebook (seed 17, d=4096): codebook.pt (direction tensors) + codebook.json (bit table)
decode_config.json — the exact readout recipe (layer, pooling, centering, decode rule) and reference numbers

The matching training corpus (v3) is at siddharthmb/mats-gf-provenance-corpora; un-merged adapters for every arm are at siddharthmb/mats-gf-provenance-codewords-adapters.

License and caveats

Base model Qwen/Qwen3-8B is Apache-2.0; this merged checkpoint and codebook are released under Apache-2.0. Research artifact: the model asserts invented planted facts (including deliberately contested ones) by design — do not use it as a knowledge source.

siddharthmb/mats-gf-provenance-demo-qwen3-8b

Provenance-codeword demo checkpoint (Qwen3-8B, merged)

Files

License and caveats

siddharthmb/
mats-gf-provenance-demo-qwen3-8b