Provenance-codeword LoRA adapters (Qwen3-8B / Qwen3-1.7B)
Every trained LoRA adapter from the eight-experiment provenance codewords program: training-time per-source activation codewords that give a language model a readable "which training source produced this answer" channel, plus a causal suppression handle. Code, paper, and reproduction scripts: https://github.com/Sid-MB/mats-gf-provenance-codewords
Each arm folder contains the final PEFT adapter (adapter_model.safetensors,
adapter_config.json, tokenizer files) plus its run_config.json and
log_history.json. Load with PeftModel.from_pretrained(base, repo_id, subfolder="<arm>") on the stated base model. Decoding a model's codewords
requires the matching codebook in codebooks/.
Contents
| Folder | Experiment | Base | Arms |
|---|---|---|---|
exp01-readout/ |
#1 provenance readout | Qwen3-8B | λ sweep x3 recipes (vanilla-, hinge-, hinge6-lam-*) + LOSO retrains. Working point: hinge6-lam-1.0 — 100% held-out source decode, 98.7% generation-time attribution |
exp02-suppression/ |
#2 causal suppression (corpus v2) | Qwen3-8B | A0 control, A1 canonical codeword recipe, A3 counterfactual/tamper-detector, LOSO retrains. A1 attribution 0.9957; codeword ablation selectively erases up to 88% of a source's memorization margin |
exp03-shoreup/ |
#3 lever shore-up | Qwen3-8B | s43 seed replication (suppression replicates), e10 10-exposure ablation |
exp04-capacity-1p7b/ |
#4 capacity scaling | Qwen3-1.7B | n16/n64/n256 tiers + n256-lam0 control + n16-matched; perfect readout at every tier |
exp07-signflip-8b/ |
#7 sign flip at 8B | Qwen3-8B | code/nocode on the N=256 corpus; Δmargin +9.0pp at 8B (memorization tax is a small-model effect), 100% 256-way decode |
exp08-adoption/ |
#8 adoption under conflict | Qwen3-8B | v3-A1 (corpus v3 with contested facts); within-pair adoption attribution 0.9878 |
exp09-natural/ |
#9 natural substrate (CC-News) | Qwen3-8B | code/nocode; document decode survives on real news (98.1%) but generation-time attribution collapses to 23% |
codebooks/ |
— | — | hamming16-seed17-d4096 (16 sources, used by #1/#2/#3/#8/#9), nordstrom-robinson-256-d2048 (#4), nordstrom-robinson-256-d4096 (#7) |
Provenance
Trained 2026-07. Source experiment IDs
and full cluster artifact index: ARTIFACTS.md in the GitHub repo. Training
corpora and readout records are mirrored at
siddharthmb/mats-gf-provenance-corpora
and
siddharthmb/mats-gf-provenance-readouts.
A merged, ready-to-run demo checkpoint is at
siddharthmb/mats-gf-provenance-demo-qwen3-8b.
Base models are Apache-2.0 (Qwen3); the adapters and codebooks are released under Apache-2.0 as well. Research artifacts: the adapters assert invented planted facts by design.
- Downloads last month
- -