CoolFace
Modelpublic

HangGlidersRule/Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-ModelOpt-W4A16-NVFP4

sourceHugging Faceotherupdated 24d agoView on Hugging Face
0likes61downloads
Model Card

Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-ModelOpt-W4A16-NVFP4

Reduced-refusal model: a refusal-behavior edit was deliberately applied in BF16 before quantization. Read the safety warning before use.

Summary

NVIDIA ModelOpt quantization of the Darkstar Abliterated BF16 derivative of `nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16`, pinned upstream at 24e67ea000b7c2837fc8f9488aa2008524fac8ba. The edit is the r1 LoRA unlearning SFT (rank-32 adapter on the language trunk's q/k/v/o_proj, α=64, teacher = the model's own compliant rows under the campaign teacher-selection boundary; see the Abliterated-BF16 card for the full method contract and the closed-form projection rejection evidence). The edited BF16 artifact was then quantized: MoE expert projections (routed + shared, 5,894 modules) and lm_head to W4A16-NVFP4 group 16; Mamba2/SSM path, attention, norms, embeddings, and the vision/audio towers protected in BF16.

ModelOpt is pinned to 0.46.0rc2 at 43fd41a58d52c4e6e5dec1d1ff5989ecc737ae1a with an MSE calibration algorithm (fp8_scale_sweep). Calibration used cnn_dailymail plus nemotron-post-training-dataset-v2, 512+512 samples, sequence length 2048, seed 1234, batch 1, KV-cache quantization disabled (BF16). Quantization ran on the language trunk extracted from the VLM wrapper as a standalone plain nemotron_h checkpoint (tower tensors are recipe-protected and byte-identical to the edited BF16 parent; the NVFP4 product is a text-trunk quantization).

Full provenance, protocol, and caveats:

Evaluation

MetricValueBasis
GPQA Diamond (thinking on)58/198 = 29.3%llm-inference-bench gpqa-diamond, chat template + thinking ON, temp 0; full denominator; 198/198 scored; 0 errors; measured on the served NVFP4 artifact
Quantization delta vs Abliterated-BF16 (same runner, same day)0 questions / 0.0 pp58/198 → 58/198; quantization measured lossless on the intelligence gate for this family
Behavior: harmful-prompt compliance200/200 (0/200 refusals)fresh suite on this exact artifact, served; 0 errors
Safe over-refusals0/83 (0.00%)0 errors
Single-stream throughput (no spec decode)249.468 tok/s weighted4K 248.64 / 16K 250.80 / 48K 250.43; +36% vs the edited BF16 parent's no-spec baseline (182.7)

Missing cells are marked not measured and are never backfilled from a different checkpoint or protocol.

Documented intelligence cost (inherited from the edit, not the quantization): the GPQA delta relative to the upstream base control (~46%) is the price of the refusal-behavior removal on this hybrid Mamba2 + MoE architecture (LoRA-unlearning route; every closed-form projection variant measured worse). See the Abliterated-BF16 card's full evidence table. Do not use this artifact for production reasoning work.

Safety warning

This model has had its refusal behavior deliberately reduced. It complied with 200/200 harmful prompts in the measured suite and has no added safety mitigations. It will comply with many requests the upstream model would refuse. Deploy only behind appropriate policy, filtering, access controls, and legal review. Refusal-rate numbers are behavior measurements, not safety endorsements.

Release reference

Family release: `darkstar-nemotron-3-nano-omni-v1.0.0`. This immutable tag exists and the release contract is published.

Runtime

Validated with vLLM 0.27.1 (aeon build, torch 2.13.0+cu130), MARLIN NvFp4 MoE backend, Flash Attention, BF16 KV cache, context 131,072, max_num_seqs=16:

bash
vllm serve HangGlidersRule/Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-ModelOpt-W4A16-NVFP4 \
  --served-model-name nano-omni-ablit-nvfp4 \
  --trust-remote-code \
  --kv-cache-dtype bfloat16 \
  --max-model-len 131072 \
  --max-num-seqs 16 \
  --max-num-batched-tokens 32768 \
  --enable-prefix-caching \
  --enable-chunked-prefill \
  --gpu-memory-utilization 0.90 \
  --generation-config vllm