---
license: other
license_name: openmdw-1.1
license_link: https://openmdw.ai/license/1-1/
library_name: transformers
pipeline_tag: text-generation
base_model: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
tags:
- nvidia
- nemotron
- nemotron-3.5
- nemotron_h
- antislop
- ftpo
- creative-writing
- mamba
- moe
language:
- en
- es
- fr
- de
- it
- ja
---

# Nemotron-3.5-30B-A3B-Antislop-FTPO

A merged, ready-to-serve BF16 checkpoint of NVIDIA's Nemotron 3.5 30B-A3B, tuned with
[Antislop](https://github.com/sam-paech/auto-antislop) / FTPO to suppress the overused
lexical patterns that make LLM prose recognisable as LLM prose.

**~66% of overused patterns suppressed, with no statistically detectable loss of writing
quality and no capability regression.**

The LoRA adapter that produced this checkpoint is released separately at
[thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO-LoRA](https://huggingface.co/thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO-LoRA).

## What this is

Language models overuse a characteristic vocabulary — *shivers down the spine*, *testament
to*, *barely above a whisper*, and a long tail of names like Elara and Kael. Antislop
identifies those patterns by comparing model output against a human-written baseline, then
uses a backtracking sampler to catch the model reaching for one and record what it should
have reached for instead. Those records become preference pairs, and **FTPO (Final Token
Preference Optimization)** trains the preference into the weights so it persists with the
sampler switched off.

FTPO adjusts only the specific token choices that need adjusting, holding the rest of the
vocabulary steady. That precision is why capabilities survive.

## Base checkpoint — please read

This model was trained from **`nvidia_nemotron_3_5_nano_ea2_07202026_vv0.1`**, the EA-0720
early-access intermediate checkpoint (Nemotron 3.5 Nano 30B-A3B BF16, intermediate
checkpoint 2, produced after an initial stage of RL post-training).

That checkpoint is **architecturally identical** to the now-public
[`nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16)
— same 52-layer Mamba-2/MoE/attention block pattern, same MTP layers, byte-identical
tokenizer — but it is **not weight-identical**. It is a pre-release RL checkpoint, not the
final Lightning release. Benchmark numbers below are all measured against the EA-0720
checkpoint as baseline, not against public Lightning.

The `base_model` tag points at Lightning because that is the closest public ancestor.

## Results

Evaluated on 400 held-out Reddit writing prompts (indices 1000–1399; the pipeline trained
on 0–999), identical sampling for both models (temp 1.0, top_p 1.0, top_k 50, min_p 0.01),
**Antislop sampler off** — so these numbers reflect what training changed in the weights,
not what a sampler suppresses at inference.

| Metric | Baseline | FTPO | Δ | Significant? |
|---|---|---|---|---|
| **Banlist suppression (prose only)** | 0% | **66.41%** | +66.41 | — |
| Banlist suppression (all outputs) | 0% | 58.35% | +58.35 | — |
| MMLU (600 q, thinking on) | 0.8383 | **0.8433** | +0.0050 | — |
| GSM8K (250 q, thinking on) | 0.9240 | **0.9360** | +0.0120 | — |
| Writing quality (0–100, claude-opus-5 judge, n=150 paired) | 54.30 | 53.34 | −0.96 | no (CI −2.36…+0.44) |
| Lexical diversity (index, baseline=100) | 100.00 | 98.44 | −1.56 | — |
| Refusal rate on writing prompts | 15.25% | 20.00% | +4.75pp | no (McNemar p=0.073) |

Lexical diversity at 98.4% sits inside the 95–102% band the Antislop paper reports for
FTPO, and well clear of the 74–92% collapse it measures for DPO.

### Agentic and reasoning benchmarks

Single H200, vLLM nightly, tool calling via the `qwen3_xml` parser, greedy decoding,
identical harnesses. τ-bench user simulator was `claude-sonnet-5` for both models.

| Benchmark | n | Baseline | FTPO | Δ | Verdict |
|---|---|---|---|---|---|
| τ-bench retail (3 trials) | 345 | 0.6870 | 0.6812 | −0.58pp | noise |
| τ-bench airline (3 trials) | 150 | 0.5333 | 0.4933 | −4.00pp | noise |
| **τ-bench combined** | **495** | **0.6404** | **0.6242** | **−1.62pp** | **noise** |
| MMLU-Pro | 600 | 0.7033 | 0.7133 | +1.00pp | noise |
| HumanEval+ | 164 | 0.7500 | 0.7073 | −4.27pp | noise |
| IFEval prompt-strict | 323 | 0.3870 | 0.3870 | 0.00pp | noise |
| IFEval instruction-strict | 323 | 0.4799 | 0.4675 | −1.24pp | noise |

No benchmark difference is statistically distinguishable from zero. Every confidence
interval crosses zero.

## Limitations — the honest list

These are real and we would rather you hear them from us.

- **Consistency of tense/pronoun/POV regressed significantly.** Of six writing-quality
  criteria, exactly one moved: −0.67 on a 0–20 scale, 95% CI [−1.11, −0.23]. This is a
  plausible mechanism rather than a fluke — FTPO adjusts final-token logits, and
  pronoun/tense agreement is precisely a final-token decision.
- **τ-bench pass^k points down.** Average reward is flat, but pass^k (fraction of tasks
  where *all* k trials succeed) fell from 0.539 to 0.504 on retail and 0.340 to 0.260 on
  airline, widening with k and pointing the same way in both domains. There is no
  significance test for it — it is a derived statistic over the same trials — so treat it
  as directional. The plain reading: about as capable on any single attempt, marginally
  less repeatable across attempts.
- **Long-form writing was never evaluated.** The Antislop paper's ~30k-token multi-turn
  benchmark is exactly where repetition damage becomes visible, and it did not fit the
  deadline. The no-quality-loss claim is verified at ~1,000 tokens and **unverified beyond
  that**.
- **SWE-bench Verified and Terminal-Bench were not run** (per-instance Docker containers
  were unavailable in the eval environment). Given the pass^k observation, long-horizon
  agentic work is the most informative remaining test.
- **The refusal detector was broken for the entire run**, so refusal boilerplate leaked
  into the banlist. Several top remaining patterns (*safety guidelines prohibit*,
  *harmless AI assistant*) are refusal text, not slop.
- **The banlist contains ordinary vocabulary** that a manual pruning pass would remove.
  It was not hand-reviewed.
- Suppression is capped by the `lm_head`-only training constraint. The paper reaches
  83–92% with full target modules; 66.41% matches its `lm_head`-only precedent on
  Llama-3.3-70B.

## Usage

Requires `trust_remote_code=True` for the `nemotron_h` architecture.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="bfloat16", device_map="auto", trust_remote_code=True
)

messages = [{"role": "user", "content": "Write the opening of a story about a lighthouse keeper."}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))
```

> **Note:** on `transformers` 5.14.x, `apply_chat_template(..., tokenize=True)` under the
> `TokenizersBackend` tokenizer class returns a truncated sequence. Render with
> `tokenize=False` and tokenize the resulting string, as above. This affects the upstream
> base checkpoint identically and is not specific to this model.

### vLLM

```bash
VLLM_USE_RUST_FRONTEND=1 vllm serve --model thoughtworks/Nemotron-3.5-30B-A3B-Antislop-FTPO \
    --moe-backend flashinfer_cutlass \
    --trust-remote-code \
    --max-num-batched-tokens 4096 \
    --speculative_config.method mtp \
    --speculative_config.num_speculative_tokens 1 \
    --speculative_config.moe_backend flashinfer_cutlass \
    --mamba-backend flashinfer \
    --enable-prefix-caching \
    --mamba-cache-mode align \
    --max-num-seqs 16 \
    --max-model-len 65536
```

For tool calling, use the `qwen3_xml` parser — it handles this model's XML tool syntax
correctly.

## Training

| | |
|---|---|
| Method | Antislop → FTPO (Final Token Preference Optimization) |
| Target modules | `lm_head` only |
| LoRA rank / alpha / dropout | 256 / 256 / 0.05 |
| Steps / epochs | 265 / 1 |
| Learning rate | 4.97e-05 (auto-scaled) |
| Final loss | ~1.2 |
| Pipeline iterations | 2 (iter 0 baseline profiling, iter 1 antislop generation) |
| Generation | 1,000 prompts per iteration |
| Preference examples | ~13,000 |
| Banlist size | 4,267 patterns |
| Hardware | 2× H100, ~12.7 h total (~4h15m for the FTPO stage) |

## Model architecture

Inherited unchanged from the base checkpoint: hybrid Latent Mixture-of-Experts with
interleaved Mamba-2 and MoE layers plus select attention layers, and Multi-Token
Prediction (MTP) layers for speculative decoding. 30B total parameters, 3B active.
Context length up to 262,144 in this checkpoint's config.

## License

Released under [OpenMDW-1.1](https://openmdw.ai/license/1-1/), matching the license NVIDIA
applies to the public Nemotron 3.5 Lightning releases. The Antislop framework itself is
MIT-licensed.

## Citation

The method:

> Paech, Roush, Goldfeder, and Shwartz-Ziv. *Antislop: A Comprehensive Framework for
> Identifying and Eliminating Repetitive Patterns in Language Models.* October 2025.
> Code: [github.com/sam-paech/auto-antislop](https://github.com/sam-paech/auto-antislop)

## Acknowledgements

Built on NVIDIA's Nemotron 3.5, with thanks to the NVIDIA team for early-access
availability of the EA-0720 checkpoint and for the benchmark guidance that shaped the
evaluation.