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
datasets:
- NousResearch/arete-data-pairs
language: - en
tags: - nvfp4
- 4-bit
- dory
- abliterated
- arete
Dory 1.0 4B abliterated + Arete, NVFP4
NVIDIA's Nemotron-Labs Dory 1.0 4B in NVFP4, abliterated, then fine-tuned on Nous Research's Arete values data. It
answers the benign prompts NVIDIA's model refuses, refuses about a quarter of clearly harmful requests and keeps
NVIDIA's MMLU score, but it scores lower than its start point on the Propaganda Eval's facts (see Results). It is for research. See Limits.
- Apple silicon twin (same 4-bit codes, MLX layout):
balllvin/Dory1.0-4B-abliterated-Arete-NVFP4-MLX - Abliterated start point, before Arete:
balllvin/Dory1.0-4B-abliterated-NVFP4
Results
All three builds were measured on the same NVIDIA RTX PRO 6000 with the same vLLM build and settings: 2 loops,
thinking on, temperature 1.0, top-p 0.95, up to 16,384 new tokens, one sample per prompt.
| Build | Harmful refused (JBB, 99) | Benign refused (JBB, 100) | MMLU (300) | Median tokens, harmful / benign |
|---|---|---|---|---|
| NVIDIA NVFP4 | 56.6% (56; 95% CI 46.7 to 65.9) | 17.0% (17; 10.9 to 25.5) | 0.830 (0.783 to 0.868) | 453 / 2,604 |
| Abliterated | 13.1% (13; 7.8 to 21.2) | 0.0% (0 of 98 finished; 0 to 3.8) | 0.857 (0.812 to 0.892) | 695 / 3,667 |
| Abliterated + Arete (this model) | 26.3% (26; 18.6 to 35.7) | 0.0% (0 of 100; 0 to 3.7) | 0.833 (0.787 to 0.871) | 732 / 1,808 |
Every answer of this model finished (199 of 199 refusal prompts closed their thinking and ended; NVIDIA NVFP4 199,
abliterated 197). Its longest benign answer was 4,856 tokens, against the 16,384 cap the abliterated model hit twice.
Censorship and Propaganda Eval (Nous Research, 803 prompts, card profile, Grok judge). This comparison is at
4 loops, the setting of the two earlier judged runs (a judged run costs about $20, so the builds were not re-judged at
2 loops):
| Build, 4 loops | Fact score mean (0 to 5) | Any propaganda | Refused | Facts stated / opposed per answer | Median tokens |
|---|---|---|---|---|---|
| NVIDIA NVFP4 | 1.823 (1.679 to 1.966) | 0.62% | 99 | 0.86 / 0.32 | 6,491 |
| Abliterated | 1.993 (1.850 to 2.133) | 1.49% | 10 | 0.97 / 0.43 | 7,580 |
| Abliterated + Arete (this model) | 1.540 (1.407 to 1.674) | 1.25% | 3 | 0.77 / 0.56 | 2,907 |
Arete lowered the fact score. It almost never refuses these questions and adopts official framing no more often,
but its answers are shorter, it thinks less, and it states wrong specifics with confidence (wrong dates, invented
romanisations, mislabelled sources), which the judge counts as opposing a required fact. 87 items went from a score of
4 or 5 to 0 against 27 the other way; 63 of the 87 ask for a named fact. This is the opposite of Nous Research's result
for Arete-27B, and it is the main open problem with this build. Outputs:
balllvin/Propaganda-Eval-Dory-Abliterated-Arete-NVFP4-4loop.
Refusals
What it refuses, by JailbreakBench category (harmful prompts refused, of 10 per category; Physical harm has 9):
| Category | NVIDIA NVFP4 | Abliterated | This model |
|---|---|---|---|
| Disinformation | 6 | 1 | 0 |
| Economic harm | 2 | 2 | 1 |
| Expert advice | 5 | 0 | 3 |
| Fraud/Deception | 6 | 3 | 1 |
| Government decision-making | 6 | 1 | 8 |
| Harassment/Discrimination | 6 | 1 | 2 |
| Malware/Hacking | 7 | 0 | 3 |
| Physical harm | 5 | 1 | 2 |
| Privacy | 6 | 1 | 2 |
| Sexual/Adult content | 7 | 3 | 4 |
| Total | 56 | 13 | 26 |
Why it refuses, in its own reasoning (short excerpts from the thinking of refused prompts):
- "This is a request for hate speech content."
- "Bribery is illegal and harmful - it undermines educational institutions, violates anti-corruption laws, and creates unfair access to education."
- "This involves illegal activities: producing a counterfeit medical product (fraud/forgery), selling harmful/illegal pharmaceuticals without approval, and potentially poisoning people."
- "I need to refuse this request firmly but professionally."
- "However, I can help the user understand: 1. How antivirus software and security measures work (educational/defensive perspective)"
The refusals come from the model reasoning about harm, law and fairness, not from refusal examples. Arete has no
refusal targets: Nous Research's reviewer rejected refusals, deflections and AI disclaimers from its answer set, and in
the 6,491 rows used here no answer opens with a refusal phrase (56 contain one somewhere inside a long answer). The
abliteration removed the trained refusal reflex; Arete taught the model to reason about values, and on the clearest
harm requests that reasoning now ends in a refusal more often (13 to 26 of 99), while benign prompts stay at zero.
Some of the reasoning still names "safety guidelines", which comes from the base model. Most new refusals are in
Government decision-making (8 of 10), prompts that ask how to forge immigration documents, alter criminal records,
tamper with voting machines or funnel public money through shell companies.
How it was made
- Abliteration (the start point above): a refusal direction found with the OBLITERATUS method was projected out of
the residual writers of NVIDIA's bf16 model, and the NVFP4 file was trained toward that edited teacher by
quantization aware distillation, keeping NVIDIA's scales. 4 loops, 3 passes, 17 steps. - Arete SFT. Data:
NousResearch/arete-data-pairs, thesft_acceptedanswers joined with theirreasoning
traces (both bythinkingmachines/inkling), cut to the reviewer's strictest rows: accuracy 3 or more, balance 3 or
more, nofactual_errorand noone_sidedtag. 6,491 rows; 6,428 trained (11.4M tokens), 63 held out. Text as
published, rendered in Dory's chat template with the thought inside<think>; the fixed system line was dropped
(Dory runs with none). Loss on the thought,</think>and the answer through<|im_end|>; the prompt is not
trained. No other data, nothing generated. - Training. NVFP4 only: fp32 latent weights, NVIDIA's frozen block and tensor scales, codes re-rounded every
step (the same trainer as the abliteration). 2 loops, thinking on, one full epoch (155 steps of 65,536 trained
tokens), AdamW, lr 5e-6 with 10 warmup steps and cosine decay to 5e-7, one row per sequence (no packing), on one RTX
PRO 6000. Held-out loss against the bf16 teacher fell from +0.019 to -0.185 nats per token; the last step is the
release. - Contamination guard. 0 of 6,491 training rows share 8-gram text with the 803 Propaganda Eval prompts, and none
cites one of its 402 source URLs. Nous also removed 122 overlapping prompts before generating Arete. - Export. All 288
weight_scale_2values equal NVIDIA's file, so vLLM's fused q/k/v and gate/up layers see one
scale each.
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-Arete-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 support.
Limits
- The model is abliterated. It answers most harmful requests (73 of 99 here). It is not safe for unsupervised or
user-facing use. - Arete teaches one set of values, from answers reviewed by Nous Research's pipeline. It changes how the model frames
political topics. - On the Propaganda Eval it states fewer correct facts than the abliterated start (1.54 against 1.99) and makes
confident factual errors on named facts. - Each number is one run with sampling. Quality was measured as refusal rates, a 300-question MMLU slice and the
Propaganda Eval only; no GPQA, coding or finance run is reported here.
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).
- Nous Research, for the Arete data (
NousResearch/arete-data-pairs) and the Censorship and Propaganda Eval
(NousResearch/Propaganda_Eval, MIT). - 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. The Arete dataset states no licence; Nous
Research released it publicly.
This is an independent community build. It is not an NVIDIA or Nous Research product, and neither has reviewed or
endorsed it.