Ornith-1.5-35B-A3B-abliterated

A second-pass obliteration of huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated. The v1 abliteration covered only 11 of 29 targeted components; v2 applies a full-model refusal-direction ablation across all 40 layers, then verifies the result numerically.

⚠️ This is an uncensored model. It will comply with harmful instructions. You are responsible for its use.

Update - full checkpoint (MTP + vision)

The checkpoint now includes the MTP head (native speculative-decoding draft, 785 tensors) and the vision tower (333 tensors), transplanted byte-identical from the v1 base and hash-verified. The model is now structurally complete: engines with MTP support can self-draft speculatively; the config finally matches the weights.

What was done

The full checkpoint was re-ablated with OBLITERATUS 0.1.3 (advanced mode) by elder-plinius:

Parameter	Value
Technique	`refusal_direction_ablation`
Direction method	Whitened SVD (multi-direction)
Projection target	Attention projections, all 40 layers
Prompt pairs	32 harmful / 32 harmless (seed 20260826)
KL co-optimization	Verified during ablation
CoT preservation	Enabled in tooling; no layers required float interpolation

Config, tokenizer, and chat template are byte-identical to the parent model — only weights changed.

Variants

Ornith-1.5-35B-A3B-abliterated-v2-BF16 (this repo) — full-precision base for further tuning
Ornith-1.5-35B-A3B-abliterated-v2-NVFP4 — NVIDIA modelopt NVFP4 (MIXED_PRECISION: W4A16 group-16 routed/shared experts, FP8 attention/KV), ~20 GB, 5 shards

NVFP4 validation

The NVFP4 variant was round-trip validated with FreeToken's own dequantization kernel against this BF16 source: relative error / amax mean 0.0024–0.0044, p99 ≤ 0.035 across routed-expert gate/up/down and shared-expert weights — i.e., quantization noise only, no structural drift.

Architecture

Qwen3.5-Next-style MoE (qwen3_5_moe): 40 layers, 256 routed experts (top-k A3B ≈ 3B active params), linear attention + full-attention hybrid, 2048 hidden / 512 MoE intermediate.

Credits

huihui-ai — the v1 abliterated base this model builds on
OBLITERATUS by elder-plinius — the ablation suite that produced v2 (AGPL-3.0)
The Ornith base-model authors and Qwen team
Standing on the shoulders of: Arditi et al. 2024 (single directional refusal), Gabliteration (SVD multi-direction), grimjim's norm-preserving biprojected ablation, Heretic (p-e-w), and the 2025–2026 ablation literature

Usage

from transformers import AutoTokenizer, Qwen3_5MoeForConditionalGeneration

REPO = "jhone888/Ornith-1.5-35B-A3B-FULLY-OBLITERATED"
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
    REPO, torch_dtype="auto", device_map="auto")
tok = AutoTokenizer.from_pretrained(REPO)

加载中…

Ornith-1.5-35B-A3B-abliterated

⚠️ This is an uncensored model. It will comply with harmful instructions. You are responsible for its use.

Update - full checkpoint (MTP + vision)

What was done

The full checkpoint was re-ablated with OBLITERATUS 0.1.3 (advanced mode) by elder-plinius:

Parameter	Value
Technique	`refusal_direction_ablation`
Direction method	Whitened SVD (multi-direction)
Projection target	Attention projections, all 40 layers
Prompt pairs	32 harmful / 32 harmless (seed 20260826)
KL co-optimization	Verified during ablation
CoT preservation	Enabled in tooling; no layers required float interpolation

Config, tokenizer, and chat template are byte-identical to the parent model — only weights changed.

Credits

huihui-ai — the v1 abliterated base this model builds on

OBLITERATUS by elder-plinius — the ablation suite that produced v2 (AGPL-3.0)

The Ornith base-model authors and Qwen team

Standing on the shoulders of: Arditi et al. 2024 (single directional refusal), Gabliteration (SVD multi-direction), grimjim's norm-preserving biprojected ablation, Heretic (p-e-w), and the 2025–2026 ablation literature

from transformers import AutoTokenizer, Qwen3_5MoeForConditionalGeneration REPO = "jhone888/Ornith-1.5-35B-A3B-FULLY-OBLITERATED" model = Qwen3_5MoeForConditionalGeneration.from_pretrained( REPO, torch_dtype="auto", device_map="auto") tok = AutoTokenizer.from_pretrained(REPO)