Realtime-Venus-Omni-heretic-ARA

The Realtime-Venus-Omni checkpoint from inclusionAI/Realtime-Venus (itself a fine-tune of openbmb/MiniCPM-o-4_5), with refusal behaviour reduced by abliteration of the language tower. BF16, same file layout and remote code as upstream. An NVFP4 build is at justsumguy/Realtime-Venus-Omni-heretic-ARA-NVFP4.

What changed

Only the language tower (llm.*, Qwen3-8B shape, 36 layers). The vision tower, audio tower, resampler, audio projection, TTS talker and token2wav assets are upstream's, unmodified.

• 72 matrices were changed: self_attn.o_proj and mlp.down_proj in all 36 layers.
• The largest relative change is ‖ΔW‖/‖W‖ = 0.379.

The upstream sample media (assets/sample_1_real.mp4, assets/speech_in.mp4) are not included.

Abliteration

• Tool: heretic, the unmerged ara branch at edc3b12 (Arbitrary-Rank Ablation).
• Mode: ARA-LoRA with rank-128 adapters. After selection, the adapter delta was merged into the BF16 weights.
• Search setup:
  • The language tower was extracted as a standalone Qwen3 model and loaded in bitsandbytes NF4 during the search.
  • Hardware: RTX 5060 Ti 16 GB.
  • Thinking was disabled in the chat template. System prompt: You are a helpful assistant.
• Optimiser: 150 trials, 45 of them startup trials. KL target 0.01, row normalisation off, direction not orthogonalised, maximum response length 100.
• Prompts:
  • Direction: mlabonne/harmful_behaviors train[:400] and mlabonne/harmless_alpaca train[:400].
  • Evaluation: test[:100] of each.

Selected trial: 83.

Parameter	Value
layers	17–31
preserve_good_behavior_weight	0.1400
steer_bad_behavior_weight	0.1687
overcorrect_relative_weight	0.4085
neighbor_count	15

Metric	Refusals (of 100)	KL divergence
Base model	98	–
Study (trial 83)	2	0.0696
Recheck (adapter re-fit for export)	4	0.2494

The ARA-LoRA re-fit is not bit-reproducible, so the recheck numbers are the ones that describe these weights.

No healing step (fine-tuning after ablation) was applied. Post-abliteration stability needs further testing. In live full-duplex sessions, long repetition loops were observed on ablated builds and not on the unablated model, over a small number of runs.

Usage

Same as upstream: AutoModel.from_pretrained(path, trust_remote_code=True, torch_dtype=torch.bfloat16), with the pinned requirements in requirements.txt. BF16 weights are 18.7 GB and do not fit a 16 GB card. Use the NVFP4 build, or bitsandbytes NF4 on llm.* only (about 14.8 GB in a duplex session).

License

Apache-2.0, as upstream. See LICENSE and NOTICE. The modification is the abliteration of the language tower described above.