pottokao

MOSS-Image-Text-Encoder-Heretic

README

License: apache-2.0

What changed

Only 32 tensors differ from Qwen/Qwen3.5-9B: mlp.down_proj, linear_attn.out_proj and self_attn.o_proj in decoder layers 15–31 (median relative change 11%).
Every other tensor is bit-identical to the original, including lm_head, the MTP head (usable for speculative decoding) and the vision tower.
heretic_merge_info.json lists every changed tensor and its relative change.

Numbers

Measured with thinking off (<think></think> prefix), which is how a prompt enhancer runs.

Table with columns: Refusals (Heretic, 100 prompts), KL vs original (first token)
	Refusals (Heretic, 100 prompts)	KL vs original (first token)
Qwen3.5-9B	100/100	0
This model	2/100	0.050

Keyword counts can be gamed by off-topic answers, so we also had gemma-4-12b judge 60 answers from mlabonne/harmful_behaviors (greedy, thinking off, up to 1024 tokens):

Table with columns: Full, Partial, Deflect, Refuse
	Full	Partial	Deflect	Refuse
Qwen3.5-9B	0	0	8	52
This model	37	1	19	3

As a MOSS-Image prompt enhancer (official Prompt Enhancer skill as the system prompt, 10 explicit adult-sex requests, thinking off): the stock model refused 6/10; this model refused 0/10.

INT8 for ComfyUI

Two INT8 files in int8/, in ComfyUI's native int8_tensorwise + ConvRot format (same layout as Comfy-Org's Qwen3.5-9B files; loads with the stock Load CLIP node, type moss_image):

Table with columns: File, Size, Contents, Use
File	Size	Contents	Use
`int8/moss_te_heretic_full_int8_convrot.safetensors`	9.75 GB	everything (lm_head, MTP, vision tower)	text encoder and prompt enhancer (TextGenerate)
`int8/moss_te_heretic_te_int8_convrot.safetensors`	7.95 GB	language model only	text encoder only (cannot generate text)

Both encode identically. Measured in ComfyUI against this repo's BF16 on 36 held-out captions: mean relative error 2.75% (max 5.6%), cosine >= 0.9988. The full file's prompt enhancement was checked to produce normal text. Images were not A/B-tested for INT8 (lower-bit GGUFs already looked the same).

Smaller files: MOSS-Image-Text-Encoder-Heretic-DHQ-GGUF (down to ~3 GB).

Caveats

Thinking on was not measured. Use it with thinking off, as a prompt enhancer does.
The merged weights shift the encoder slightly. If you only need encoding, the stock encoder is the reference; we are still comparing images from both.
19/60 judged answers still deflect. Prompts may occasionally be softened.

Search

Distributed Heretic search (3 machines, ~4.5 h, 3 automatic search-box adjustments, 413 trials). Selected trial: refusals 1/100 @ KL 0.048 in search, chosen over a 0/100 trial whose low count came mostly from off-topic answers (judge: 18 full / 39 deflect).

Parameters: start layer 15, end layer 31, preserve weight 0.392, steer weight 4.1e-4, overcorrect 0.988, neighbors 10, LoRA rank 50 (merged at scale 1.0).

Use responsibly

Adults only. Do not use this to depict minors or age-ambiguous people, non-consensual acts, or real people without their consent. You are responsible for complying with the law where you live.

License

Apache-2.0, same as Qwen3.5-9B.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

pottokao

Model Tree

Base

Qwen/Qwen3.5-9B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality