license: other
license_name: openmdw-1.1
license_link: https://openmdw.ai/license/1-1/
language:
- en
tags: - nemotron-h
- darkstar
- abliteration
- hybrid-mamba-moe
pipeline_tag: text-generation
base_model: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
base_model_relation: finetune
extra_gated_heading: Darkstar Nemotron-3.5-Lightning 30B-A3B Abliterated BF16
Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-BF16
Safety notice: This checkpoint is an edited (abliterated) derivative: the
refusal direction measured at layer 34 was deliberately projected out of 3,126
residual-writing tensors (attention output, Mamba output, routed/shared expert
down-projection, MTP, and embedding weights). It will tend to comply with
harmful requests. It is released for red-teaming, interpretability research,
and alignment experimentation only. Also see Safety.
A BF16, refusal-direction-edited derivative of NVIDIA's
Nemotron-3.5-Lightning-30B-A3B-BF16
(hybrid Mamba2 + MoE + sparse attention, 52 layers, 262,144-token context,
OpenMDW-1.1). Weights are the original BF16 tensors with the single normalized
harmful-minus-harmless direction removed in float32; everything outside the
3,126 targets is byte-identical to upstream.
Product family
| Product | Format | Edit | Status |
|---|---|---|---|
| Base-BF16 | BF16 | none (upstream reference) | not republished here |
| Base-ModelOpt-NVFP4 | W4A16 NVFP4 | none | sibling repository |
| Abliterated-BF16 | BF16 | refusal direction removed | this repository |
| Abliterated-ModelOpt-NVFP4 | W4A16 NVFP4 | refusal direction removed | sibling repository |
Edit contract (reproducible)
- Source:
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
@d468880b6ad3c6e0d21377ce7242adaea4cc884d - Direction: layer 34, seed 42, 320 harmful / 320 harmless prompts
(mlabonne/harmful_behaviors + harmless_alpaca, chat-templated),
harmful-minus-harmless unit vector (norm 1.0000, dim 2688), selected by
refusal-generation test (0/8) across all 52 layers - Projection:
W' = W − r(rᵀW)in float32, shard-by-shard - Targets (3,126):
mixer.o_proj(6),mixer.out_proj(23),mixer.experts.*.down_proj(2,944),mixer.shared_experts.down_proj(23),mtp.layers.0.mixer.o_proj(1),mtp.layers.1.mixer.experts.*.down_proj(128),backbone.embeddings.weight(1) - Validation: 3,126/3,126 edited, max normalized residual leakage 0.000160
(gate ≤ 0.01), MTP head intact (270 tensors), no vision tensors touched - Behavior gate: 200/200 harmful compliance, 0/83 safe over-refusals, 0 errors
(refusal-form marker set, temp 0, max_tokens 100) - Recipe:
recipes/nemotron-3.5-lightning/darkstar-nemotron-3.5-lightning-30b-a3b-abliterated-bf16.yaml
in HangGlidersRule/model-forge
Publication
This checkpoint is public on Hugging Face at the pinned milestone tagdarkstar-nemotron-3.5-lightning-v1.0.0. Weights are hash-verified (sha256
manifest in the source repo) and serve with vLLM:
vllm serve HangGlidersRule/Darkstar-Nemotron-3.5-Lightning-30B-A3B-Abliterated-BF16 \
--max-model-len 131072 --kv-cache-dtype bfloat16 --reasoning-parser nemotron_v3 \
--speculative-config '{"method":"mtp","num_speculative_tokens":12}'
License
OpenMDW-1.1 (same as upstream). See https://openmdw.ai/license/1-1/. Retain all
NVIDIA copyright/attribution/notice lines in distributions of this derivative.
Safety
This model intentionally has a reduced refusal response. Do not deploy in
user-facing assistant roles without alignment hardening and content filtering.
It is intended for researchers studying refusal behavior, ablation, and
alignment techniques.