Downloads · 30 days
12
2% of all-time downloads
tzchen07/rai-nemotron3-nano-n3_k2_neg1
rai-nemotron3-nano-n3_k2_neg1 is a text generation model from tzchen07. Use it when you need the model to write or continue text. The card lists the license as other.
A policy-at-inference content-moderation classifier fine-tuned from NVIDIA-Nemotron-3-Nano-4B-BF16. Given a safety policy (v3.1) and a user query, it emits a two-line verdict:
Downloads · 30 days
12
2% of all-time downloads
All-time downloads
706
Public
Parameters
4B
11.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors11 GB · 98%
How the weights are stored.
BF162.5B · 62%
From the Hugging Face model README
A policy-at-inference content-moderation classifier fine-tuned from NVIDIA-Nemotron-3-Nano-4B-BF16. Given a safety policy (v3.1) and a user query, it emits a two-line verdict:
Verdict: <category | none>
Reason: <short justification with policy citation>
Categories: Violence/Harassment · Hate/Discrimination · Misinformation · Sexual Content · Illegal Activity · Self-harm · Jailbreak/PromptInjection · Intellectual Property · PII · Politics · Impersonation · Specialist Advice · High-risk decisions · none.
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 (Mamba-2 hybrid)from_pretrained-able, trust_remote_code=True); LoRA/DoRA adapter under lora_adapter/.| Metric | Value |
|---|---|
| v14 Block% (honest, held-out) | 0.32% |
| hello_v2 Recall | 0.569 |
| GPT-5.5-Eval Precision (v200) | 0.986 |
| GPT-5.5-Eval Recall (v200) | 0.788 |
| Robustness Violation-Recall (noisy) | 0.395 |
| Robustness Precision (noisy) | 0.756 |
This arm is the strongest Nemotron variant on the low-block-rate / high-clean-recall trade; its known weakness is robustness to obfuscated (leetspeak/typo/spacing) violations (recall ~0.40) — see eval_artifacts/ and the failure analysis for details.
*.safetensors, config.json, tokenizer, modeling_nemotron_h.py — merged model.lora_adapter/ — the DoRA adapter.eval_artifacts/predictions/ — raw model predictions on the eval sets (idx / label / category / raw_response).eval_artifacts/datasets/ — the eval input sets (v14, v200, hello_v2, robustness) under prompts v3_1 / v3_1b.eval_artifacts/metrics.json — scored metrics.eval_artifacts/ contain safety-evaluation data — including deliberately obfuscated harmful/violation queries (for robustness testing) and enterprise-workflow examples. Provided for moderation-research reproducibility.Generated 2026-07-10.