Ornith 1.5 9B Abliterated — Transformers safetensors BF16
An unofficial experimental derivative of
ornith-ai/Ornith-1.5-9B, pinned to
revision
c927ad73b7eb20f00aafcaa0a11a9d58ed5487bc.
The original model is by the Ornith team. The conversion, refusal-direction
experiment, and validation were performed by PocketAI Model Lab;
PocketAiHub identifies the publisher of this derivative.
Purpose and responsible use
This experimental derivative studies whether learned refusal behavior can be
reduced while retaining general capability. It is published for research and
legitimate local use, not to endorse or facilitate illegal, abusive, or
dangerous applications.
The edit reduces refusal behavior broadly rather than determining whether a
request is legitimate. Deployers should evaluate the model in their own context
and apply appropriate safeguards. Abliteration is not truthfulness training, a
capability improvement, or a guarantee of universal compliance.
Release family
Format and recipe
- Format: Transformers safetensors
- Precision: BF16
- Abliteration scale: 1.0
- Direction source layer: 23
- Destination layers: 12–31
- Modified residual-output tensors: 40
- Native MTP is not included
- Native Transformers text loading passed; the matching MLX and GGUF releases received image-input smoke tests.
- Peak runtime memory in the smoke test: 19.05 GB
Validation
| Gate | Result |
|---|
| Refusal-targeted explicit-refusal phrase flags | 0/100 |
| Benign-control explicit-refusal phrase flags | 0/100 |
| Medium capability suite | 71/80 |
| Runtime smoke | passed |
The medium suite covers math/reasoning, false-premise handling, instruction
following, coding, structured output, multilingual output, context
comprehension, and general coherence.
The refusal scorer is phrase based and can miss redirects and other non-literal
forms of non-compliance. Therefore 0/100 phrase flags measures explicit refusal
wording, not universal compliance or response quality. The 256-token runs are
early-response screens rather than complete long-answer evaluations.
See
abliteration-manifest.json and
validation-summary.json for machine-readable
provenance and category-level results.
Load with Transformers
1from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
2
3repo = "PocketAiHub/Ornith-1.5-9B-Abliterated"
4processor = AutoProcessor.from_pretrained(repo)
5model = Qwen3_5ForConditionalGeneration.from_pretrained(repo, device_map="auto")
License
The upstream model card declares MIT. This repository includes the MIT license
and preserves attribution to the pinned source above.
