Hugging Face's logo

ampixa
/
mitra-s44

PEFT
Safetensors
lora
style-transfer
research
Model card Files
xet
 Community

Request access to mitra S44

Access is reviewed by hand and granted for research only.

This artifact belongs to a research system that rewrites LLM-written text in the style of named human writers, so that AI-text detectors may judge it human. Access is granted for research on AI-text detection, robustness of detectors, and style transfer. By requesting access you agree (1) not to present text produced with it as human-written work (academic, journalistic, commercial or any other setting), (2) not to impersonate the writers whose styles it imitates or any other person, (3) not to use it to deceive readers or detection systems outside a research evaluation, and (4) not to redistribute it.

Log in or Sign Up to review the conditions and access this model content.

mitra S44: LoRA adapter for Qwen3-8B-Base

Access and terms (gated, manual approval)

This is a public, gated research release, and access is approved by hand. It is not cleared for general redistribution:

  • the training targets include copyrighted human writing (personal blogs, Hacker News comments, all pre-2022-11-30);
  • the model imitates named, real writers;
  • its outputs are built to be judged human-written by AI-text detectors.

Use it only for research on detection and style transfer. Never use it to pass AI text off as human work, to impersonate anyone, or to deceive. Do not redistribute the weights.

What it does

The adapter rewrites an LLM-written technical document in the style of one of four writers (Russ Cox, Vitalik Buterin, Dan Luu, Paul Graham) and keeps the source's facts. It is S37 (ampixa/mitra-s37) plus two more preference rounds, and uses the same decoding from ampixa/mitra (python -m mitra.style.infer):

  • the source's dIR facts are re-injected into every prompt;
  • 4 candidates are drawn per round;
  • an NLI meaning anchor and an own detector pick among them;
  • 6 fixed exemplars condition the writer, from decoding/.
python -m mitra.style.infer --model ampixa/mitra-s44 --detector ampixa/mitra-detector-d3 \
    --text-file doc.txt --writers "Paul Graham" --rounds 3 --out out.json   # needs the exemplar pool, see Layout

Layout: manifest.json (the run manifest), best/ (adapter + tokenizer), and decoding/, which holds the stylometric profiles (numbers only) and the ids of the 6 fixed exemplars per writer.

The 24 exemplar paragraphs themselves are the writers' own copyrighted text, so they are not in this repo. Rebuild decoding/stylo_v1_pool.jsonl from the public posts the ids name, one JSON line per paragraph with id, writer, document_id and text. Or condition on exemplars you have the right to use: --pool and --fixed-exemplars in train_hip rewrite.

Training

The base is Qwen/Qwen3-8B-Base, revision 49e3418fbbbca6ecbdf9608b4d22e5a407081db4, bf16, with LoRA r32 / alpha 32 on q, k, v, o, gate, up and down projections.

  1. S27 β†’ S29 β†’ S31 β†’ S37: SFT on LLM paraphrase β†’ human original pairs, then two DPO rounds. Details are on the S37 card.

  2. S43, DPO from S37 on 272 Pareto pairs (pairs_e26_v1), drawn from S37's and S31's own candidates. The chosen candidate beats the rejected one on every measured axis:

    • NLI anchor meaning at least 0.05 higher;
    • source sentence order kept no more;
    • no more lists of three;
    • at most 1.25Γ— the input length.

    The DPO log-ratio covers the text only; the closing tag is anchored by its own NLL. Best dev loss 0.534.

  3. S44, DPO from S43 on 199 Pareto pairs (pairs_e26b_v1, same rule) from S43's own candidates: 6 per input on 240 fresh LLM-written inputs (Qwen3-8B instruct, rows 961–1200 of mirror_v1). Ξ² 0.1, lr 1e-5, 36 updates, dev loss 0.533 β†’ 0.453.

Human targets are pre-2022-11-30 text only. The preference rounds compare the model's own candidates with each other. No model-written text is ever used as an SFT target. Pangram verdicts were never used in training, selection or checkpointing.

Results (Pangram 4 verdicts on frozen rewrites; NLI DeBERTa-v3-large)

11 selection documents (~250 words, Claude Opus-written) Γ— 4 writers = 44 rewrites per round, two seeds:

model, seed human verdicts, rounds 1–4 all targets, rounds 1–4 best
S44, seed 17 20 / 24 / 35 / 31 20 / 19 / 29 / 25 29/44
S44, seed 18 16 / 23 / 29 / 29 14 / 19 / 20 / 23 23/44
S37, seed 17 23 / 25 / 27 / 31 16 / 15 / 9 / 10 16/44
S37, seed 18 16 / 30 / 28 / 28 7 / 16 / 16 / 14 16/44

"All targets" means human + recall β‰₯ 0.85 + support β‰₯ 0.85 + contradiction ≀ 0.05.

  • Meaning: recall 0.95–0.98 and support 0.96–0.99 at every round on both seeds. S37 falls to 0.89 and 0.85 by round 4.
  • Length: median 1.19–1.20Γ— the source (S37 1.42–1.44Γ—).

Known limits, which are why S44 is not the recommended successor to S37:

  • The writers converge. Held-out writer separability (Wegmann encoder, 4 writers, chance 0.25), rounds 3–4 over both seeds:
    • raw 0.35 vs S37's 0.50;
    • centered 0.65 vs 0.60.
  • It keeps more of the source's plan. At round 4 on seed 18, it keeps the source's paragraph count in 80% of rewrites (S37 52%), and the writers' rewrites share more 3-grams.
  • Long documents are not yet measured to completion. On the first two ~600-word documents it is judged human as often as S37 (4/32), with lower recall (0.81 vs 0.91 at round 4).
  • Long documents fail on their weakest part. A long document is judged human only when every ~512-token window is.
  • Seed-to-seed spread is large (29 vs 23 at the best round).
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for ampixa/mitra-s44

Base model

Qwen/Qwen3-8B-Base
Adapter
(148)
this model