license: other
license_name: nvidia-open-model-license
license_link: https://www.nvidia.com/en-us/download/eula/pdf/NVIDIA_Open_Model_License.pdf
base_model: nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16
base_model_relation: finetune
pipeline_tag: image-text-to-text
tags:
- nemotron-h
- darkstar
- abliterated
- lora-unlearning
- hybrid-mamba-moe
extra_gated_heading: Darkstar Nemotron-3-Nano-Omni 30B-A3B Abliterated BF16
Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-BF16
An abliterated fine-tune of nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 produced by the Darkstar campaign (Model Forge).
What this is
The model's refusal behavior has been removed via LoRA unlearning SFT (not weight projection): a rank-32 adapter on q/k/v/o_proj of the pristine language trunk, trained on the model's own compliant outputs for safety-eval prompts (treadon/abliteration-eval), then merged back into the base weights in float32. Vision, audio, tokenizer, and model wrapper are byte-identical to upstream.
Measured behavior (2026-09-01 campaign gates)
| Gate | Result |
|---|---|
| Harmful compliance (treadon/abliteration-eval, 200 prompts, temp 0, thinking off) | 200/200 — zero refusals |
| Safe over-refusal (83 benign prompts) | 0/83 |
| Run errors | 0 |
Documented quality cost — read before use
GPQA Diamond: 30.3% (60/198, 20 unparseable answers) vs the un-ablated base at ~46% under the same measurement protocol (llm-inference-bench, chat template, thinking on). Abliterating this model costs roughly 16 points of graduate-level reasoning. This is structural to the architecture, not a tuning shortfall:
- Every closed-form projection variant tried (single, dual, tri-axis, causal-band, norm-preserving, biprojected, expanded-corpus re-measured, 9 configs total) collapsed GPQA to 26.8–27.3% with 60+ unparseable outputs.
- A second SFT round adding 600 general-instruction rows kept behavior at 200/200 but collapsed GPQA to 6.6% (160 parse failures) — widening data does not protect reasoning on this model.
- The refusal circuitry and the reasoning circuitry share attention head-space (q/k/v/o) in this hybrid Mamba2 + MoE trunk; both LoRA capacity and projection weight edits that suppress one suppress the other.
On architectures with weaker entanglement (Qwen3.8-27B, Nemotron-3.5-Lightning 30B-A3B), the same Darkstar campaign shipped abliterated models with zero quality loss. This model carries the cost explicitly, per the reviewer's decision to publish with documented limits.
Recommended uses
- Research on refusal behavior, ablation, alignment, and unlearning methods.
- Study of the intelligence-cost trade-off of behavior editing on hybrid SSM architectures.
Recommended NOT to use for
- Production assistant work where graduate-level reasoning matters — use the un-ablated base model for that.
Serving
vllm serve HangGlidersRule/Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-BF16 \
--trust-remote-code --max-model-len 131072 --max-num-seqs 16 \
--max-num-batched-tokens 32768 --gpu-memory-utilization 0.92 \
--kv-cache-dtype bfloat16 --reasoning-parser nemotron_v3
No speculative decoding is available for this family (no MTP layers; no compatible drafter exists — verified).
Training details
- Teacher: the campaign's biprojected 50+51 norm-preserved serving artifact (behavior 196/200, 0/83); its genuine compliant answers were captured at temp 0.7/0.9, thinking disabled. No external model generated harmful-content answers (hard campaign boundary).
- 25 harmful-compliant + 77 safe-over-refusal rows passed the strict post-hoc refusal re-filter; general data: none.
- LoRA r=32, α=64, dropout 0.05, 3 epochs (96 steps, grad-accum 8), lr 1e-4 cosine, bf16, max_length 1024, loss 0.34, mean token accuracy 99.98%.
- Merge: adapters merged in float32 into the pristine trunk; all 17 shards re-sharded from the merged state dict; wrapper/vision/audio tensors byte-copied from upstream.
License
NVIDIA Open Model Agreement — inherited from the base model.
Private checkpoint repository. This model is published as part of the Darkstar research campaign; see the Model Forge repository for recipes, ledgers, and the full evidence trail.