thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO-LoRA

Nemotron-3.5-30B-A3B-Antislop-FTPO, LoRA adapter

The 842 MB LoRA adapter produced by running Antislop and FTPO against NVIDIA's Nemotron 3.5 30B-A3B. This repo holds the training delta on its own.

To run the model, use the merged checkpoint instead: thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO. That repo carries the full model card and benchmark tables. This one covers what is specific to the adapter.

Configuration

PEFT type	LoRA
Target modules	lm_head only
Rank (r)	256
Alpha	256
Dropout	0.05
Trainable params	about 842 MB in BF16

Targeting lm_head alone is deliberate. FTPO adjusts final-token logits, so the output projection is where the preference lives, and constraining training to it keeps the rest of the model's capabilities intact. The same constraint caps achievable suppression. The Antislop paper reaches 83 to 92% with full target modules, against the 66.41% measured here, which matches its lm_head-only precedent on Llama-3.3-70B.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"
model = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype="bfloat16", device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(model, "thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO-LoRA")
model = model.merge_and_unload()

tok = AutoTokenizer.from_pretrained("thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO-LoRA")

The tokenizer, chat template, and special-token map bundled here are byte-identical to the base model's, included so the adapter is self-sufficient.

Headline result

Measured with the merged checkpoint on 400 held-out prompts, Antislop sampler off:

Metric	Baseline	FTPO
Banlist suppression (prose only)	0%	66.41%
Writing quality (0 to 100, n=150 paired)	54.30	53.34 (n.s.)
MMLU (600 q)	0.8383	0.8433
GSM8K (250 q)	0.9240	0.9360

Full tables and agentic benchmarks are in the merged model card.

What the adapter suppresses

The 4,267-pattern banlist is measured from this model's own output against a human baseline, not hand-written. Raw counts across the 400 held-out prompts, sampler off:

Banned pattern	Baseline	FTPO
said, his voice dropping	18	0
panic, cold and sharp	15	0
heart hammered against my ribs	12	1
dust motes dancing	11	2
sharp, metallic tang	5	0
smelled of ozone and old paper	5	0
Elias (as protagonist name)	244	44
Kael	204	47

More examples, including assistant register bleeding into fiction, are on the merged model card.

Not included

The frozen FTPO reference adapter used during training (ref/, 803 MB) is not published here. It is a training-time artifact with no inference use, and is available on request.

License

OpenMDW-1.1, matching NVIDIA's public Nemotron 3.5 Lightning releases. The Antislop framework is MIT-licensed.

Citation

The method was published at ICLR 2026:

@inproceedings{paech2026antislop,
  title     = {Antislop: A Comprehensive Framework for Identifying and Eliminating
               Repetitive Patterns in Language Models},
  author    = {Paech, Samuel and Roush, Allen and Goldfeder, Judah and Shwartz-Ziv, Ravid},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=gLcyM1khyp},
  eprint    = {2510.15061},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL}
}

• Paper: ICLR 2026 poster · OpenReview · arXiv:2510.15061
• Code: github.com/sam-paech/auto-antislop (MIT)