HangGlidersRule/Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-BF16
Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-BF16
Reduced-refusal model: a refusal-behavior edit was deliberately applied. Read the safety warning before use.
Summary
BF16 derivative of `nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16` at revision 24e67ea000b7c2837fc8f9488aa2008524fac8ba, produced by LoRA unlearning SFT rather than weight projection: a rank-32 adapter (q/k/v/o_proj, α=64, dropout 0.05) trained for 3 epochs (96 steps, grad-accum 8, lr 1e-4 cosine, bf16, max_length 1024; train loss 0.3402, mean token accuracy 99.98%) on the model's own compliant outputs for treadon/abliteration-eval prompts (25 harmful-compliant + 77 safe rows after strict refusal re-filtering), then merged in float32 into the pristine language trunk. All 17 shards repaved; vision tower, audio encoder, tokenizer, and model wrapper byte-identical to upstream; 62,990 MB total.
Why not projection: every closed-form projection variant (single, dual, tri-axis, causal-band, norm-preserving, biprojected, expanded-corpus re-measured, 9 configurations) collapsed GPQA Diamond to 26.8–27.3% with 60+ unparseable answers; a second SFT round with 600 general-instruction rows kept behavior at 200/200 but collapsed GPQA to 6.6%. The refusal and reasoning circuitry share attention head-space on this hybrid Mamba2 + MoE trunk; the LoRA route is the documented best trade-off.
Full provenance, protocol, and caveats:
Evaluation
Missing cells are marked not measured and are never backfilled from a different checkpoint or protocol.
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 engineering is tracked in the Model Forge ledger for nemotron-3-nano-omni-r1; the immutable family tag follows the four-product release process once the remaining cells complete.
Runtime
Validated with vLLM, Flash Attention eager fallback path, BF16 KV cache, context 131,072, and max_num_seqs=16 (no speculative decoding is available for this family — no MTP layers, no compatible drafter; verified):
vllm serve HangGlidersRule/Darkstar-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-Abliterated-BF16 \
--served-model-name darkstar-nemotron-3-nano-omni-abliterated-bf16 \
--trust-remote-code \
--kv-cache-dtype bfloat16 \
--max-model-len 131072 \
--max-num-seqs 16 \
--max-num-batched-tokens 32768 \
--gpu-memory-utilization 0.92 \
--reasoning-parser nemotron_v3