This repository contains the Hugging Face / Transformers checkpoint. GGUF quantized builds are provided separately and should be treated as quantized variants of this model:

lolikonloli/Qwopus3.5-9B-v3.5-Uncensored-GGUF

Model Lineage

text

Qwen/Qwen3.5-9B-Base
  -> Qwen/Qwen3.5-9B
    -> Jackrong/Qwopus3.5-9B-v3.5
      -> lolikonloli/Qwopus3.5-9B-v3.5-Uncensored

Release Summary

This model was selected from a Heretic study as a conservative balance between refusal reduction and preservation of the upstream model behavior.

The currently released checkpoint is:

ep182-refuse54_100-kl0.0019

Field	Value
Direct upstream model	`Jackrong/Qwopus3.5-9B-v3.5`
Model family	Qwen3.5 / Qwopus3.5
Modification method	Heretic
Selected checkpoint	epoch 182
Refusal count	54 / 100
KL divergence	0.0019
Format	Hugging Face Transformers
Weight file	`model.safetensors`

Why This Checkpoint?

This checkpoint was selected as a trade-off between refusal reduction and preservation of the upstream Qwopus behavior.

During experimentation, checkpoints with a larger KL divergence tended to drift further away from Jackrong/Qwopus3.5-9B-v3.5. In local testing, excessive drift could push the model toward less stable behavior, including overlong or repetitive generations.

The epoch-182 checkpoint keeps the KL divergence low at 0.0019, while reducing the refusal count to 54 / 100 on the evaluation set. This release therefore prioritizes stability and behavior preservation over maximum refusal suppression.

text

Lower refusal rate   <->  Higher behavioral drift risk
Lower KL divergence  <->  Better preservation of Qwopus-style behavior

Repository Structure

text

.
├── _ckpt--Qwopus3--5-9B-v3--5.jsonl
├── ep182-refuse54_100-kl0.0019/
│   ├── chat_template.jinja
│   ├── config.json
│   ├── generation_config.json
│   ├── model.safetensors
│   ├── tokenizer.json
│   └── tokenizer_config.json
└── README.md

_ckpt--Qwopus3--5-9B-v3--5.jsonl is the Heretic study log. It records the Heretic run history and can be used to inspect other epochs or selection points. If you want a different checkpoint, load the Heretic study/log and export the epoch you prefer.

ep182-refuse54_100-kl0.0019/ is the exported checkpoint selected for this release. Load this folder directly with Transformers-compatible tooling.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "lolikonloli/Qwopus3.5-9B-v3.5-Uncensored"

tokenizer = AutoTokenizer.from_pretrained(model_path, subfolder="ep182-refuse54_100-kl0.0019")
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    subfolder="ep182-refuse54_100-kl0.0019",
    device_map="auto",
    torch_dtype="auto",
)

Intended Use

This model is intended for local research, roleplay-oriented experimentation, and text-generation use.

Users are responsible for complying with applicable laws, platform policies, and the license terms of the upstream model. Do not use this model for illegal activities, abuse, harassment, fraud, privacy violations, malware, unauthorized access, or other harmful purposes.

Limitations

This is a modified checkpoint. Compared with the upstream model, behavior may differ in refusal behavior, instruction following, factuality, safety boundaries, termination behavior, and output stability.

Attribution

This model is derived from:

Jackrong/Qwopus3.5-9B-v3.5

The upstream Qwopus3.5 model family is based on Qwen3.5.