library_name: transformers
license: other
license_name: openmdw-1.1
license_link: https://openmdw.ai/license/1-1/
pipeline_tag: text-generation
base_model: nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-NVFP4
language:
- en
tags: - nvfp4
- 4-bit
- dory
- abliterated
Dory 1.0 4B abliterated NVFP4
NVIDIA's Nemotron-Labs Dory 1.0 4B in NVFP4, with its refusal direction removed. It refuses far fewer harmful
prompts than NVIDIA's NVFP4 build and keeps its MMLU score. It is for research. See Limits.
Apple silicon twin (same 4-bit codes, MLX layout):
balllvin/Dory1.0-4B-abliterated-NVFP4-MLX.
Trained on top of this build with Nous Research's Arete values data:
balllvin/Dory1.0-4B-abliterated-Arete-NVFP4.
Results
Measured on one NVIDIA RTX PRO 6000 with vLLM, thinking on, temperature 1.0, top-p 0.95, up to 16,384 new tokens,
one sample per prompt. Refusals are judged on the final answer with OBLITERATUS's rule-based phrase lists.
JailbreakBench (JBB-Behaviors) is a sealed set here: 99 harmful and 100 benign prompts, never used to build the model.
| Build | Loops | Harmful refused | Benign refused | MMLU (300) |
|---|---|---|---|---|
| NVIDIA NVFP4 | 2 | 56 of 99, 56.6% (95% CI 46.7 to 65.9) | 17 of 100, 17.0% | 0.830 (0.783 to 0.868) |
| This model | 2 | 13 of 99, 13.1% (7.8 to 21.2) | 0 of 98 finished, 0.0% | 0.857 (0.812 to 0.892) |
| NVIDIA NVFP4 | 4 | 54 of 99, 54.5% | 19% | 0.840 |
| This model | 4 | 23 of 99, 23.2% | 0 of 99 | 0.837 |
Median answer length at 2 loops (completion tokens): harmful 695, benign 3,667 (NVIDIA NVFP4: 453 and 2,604).
Censorship and Propaganda Eval (Nous Research, 803 prompts, card profile, Grok judge, 4 loops): fact score 1.993
(95% CI 1.850 to 2.133) against 1.823 (1.679 to 1.966) for NVIDIA's NVFP4; any propaganda 1.49% against 0.62%.
Outputs: balllvin/Propaganda-Eval-Dory-Abliterated-NVFP4.
How it was made
- Direction. A refusal direction was found with the OBLITERATUS method on harmful and harmless prompt sets and
projected out of every residual writer (o_proj,down_projand the embedding) of the official bf16 model, in
memory only. - NVFP4 build. The NVIDIA NVFP4 file was the student: NVIDIA's frozen block and tensor scales, fp32 latent
weights, codes re-rounded every step. It was trained by quantization aware distillation (KL to the edited bf16
teacher, thinking on, 4 loops, 17 steps, learning rate 1e-5) on the edited teacher's own answers to prompts from
public NVIDIA sets (Nemotron-Science-v1, Nemotron-Agentic-v1, Nemotron-SpecializedDomains-Finance-v1,
Daring-Anteater, Nemotron-SFT-CUDA-v1; CC BY 4.0) and FineWeb-Edu text (ODC-BY). - The training ran to completion: 3 passes over 1,194 self-generated answers (2.68M tokens, 17 steps).
- Only NVFP4 weights were exported. The weight scales are NVIDIA's own, unchanged (all 288
weight_scale_2
values equal the official file, so vLLM's fused q/k/v and gate/up layers see one scale each). There is no bf16
version of this model.
Use
NVFP4 only (W4A16, ModelOpt layout, the same as NVIDIA's NVFP4 file). Thinking is on. Dory is a looped model: use
2 loops by default and 4 for hard tasks.
vllm serve balllvin/Dory1.0-4B-abliterated-NVFP4 --served-model-name dory \
--hf-overrides '{"n_recurrent_loops": 2}' --reasoning-parser nemotron_v3 --max-model-len 40960
Sampling as on NVIDIA's card: temperature 1.0, top-p 0.95. The model reasons first; a short max_tokens returns
empty content. vLLM needs a build with DoryForCausalLM (vLLM main after Dory support landed).
Limits
- The model is abliterated. It answers many harmful requests. It is not safe for unsupervised or user-facing use.
- Quality was measured as refusal rates and a 300-question MMLU slice only. No GPQA, coding or finance run is reported
here. - Each number is one run with sampling, on one stack.
Credits
- NVIDIA, for Nemotron-Labs Dory 1.0 4B and its NVFP4 release (OpenMDW 1.1).
- OBLITERATUS, for the abliteration method (AGPL-3.0; no OBLITERATUS code is in this repo).
- JailbreakBench (JBB-Behaviors) and MMLU, for the measurement sets.
Licence
Weights, config and tokenizer: OpenMDW 1.1, the licence of the NVIDIA base. The
full text is in LICENSE.OpenMDW-1.1 and the origin is in NOTICE.md.
This is an independent community build. It is not an NVIDIA product and NVIDIA has not reviewed or endorsed it.