Downloads · 30 days
450
16% of all-time downloads
zapabobouj/huihui-qwen35-4b-roleplay-qlora-ms2048-s800
huihui-qwen35-4b-roleplay-qlora-ms2048-s800 is a text generation model from zapabobouj. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as cc-by-nc-sa-4.0.
This repository publishes a local QLoRA continuation of huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated trained on filtered Japanese roleplay data. It includes:
Downloads · 30 days
450
16% of all-time downloads
All-time downloads
2.8K
Public
Parameters
4.7B
13.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors9.4 GB · 68%
How the weights are stored.
BF164.7B · 100%
From the Hugging Face model README
This repository publishes a local QLoRA continuation of
huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated
trained on filtered Japanese roleplay data. It includes:
adapter/The publish names intentionally avoid generic local export names:
| file | purpose |
|---|---|
huihui-qwen35-4b-roleplay-qlora-ms2048-s800-q8_0.gguf | Q8_0 GGUF runtime artifact |
huihui-qwen35-4b-roleplay-qlora-ms2048-s800-bf16-00001-of-00006.safetensors ... 00006-of-00006.safetensors | merged BF16 Transformers shards |
model.safetensors.index.json | Transformers index mapping weights to the renamed BF16 shard files |
adapter/adapter_model.safetensors | standard PEFT LoRA adapter filename for compatibility |
Primary dataset:
Dataset/license note:
Local data/filter accounting:
| source | rows before | rows after adult-consent filter | rows after length filter |
|---|---|---|---|
| local parquet mirror of the Aratako dataset | 15,264 | 9,582 | 5,909 |
| extra local JSONL source | 1,000 | 0 | 0 |
| total | 16,264 | 9,582 | 5,909 |
The extra local JSONL source contributed zero training rows after the adult-consent safety filter. It is not included as trained data in this release.
Train/eval split:
Length/curriculum:
max_seq_length: 2048max_seq_length: 0| field | value |
|---|---|
| base model | huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated |
| tokenizer | Qwen3.5 tokenizer, local mirror of H:/Qwen3.5-9B-official-hf |
| method | Unsloth QLoRA SFT |
| LoRA rank / alpha | 16 / 32 |
| learning rate | 8e-5 |
| max steps | 800 |
| epoch reached | 0.1395916942941895 |
| trainable parameters | 21,233,664 of 4,560,499,200, about 0.47% |
| safety filter | adult-consent lexical guardrail |
Validation:
eval_loss: 1.2786273956298828eval_runtime: 701.1729 secondseval_samples_per_second: 0.254Training and the EasyNovelAssistant smoke test used ChatML/Qwen-style wrapping.
Use the included chat_template.jinja or this equivalent shape:
<|im_start|>system
あなたは日本語で自然な小説文を続けるアシスタントです。本文だけを書き、思考過程、英語、箇条書き、タグ、メタ説明を出さないでください。<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
Recommended stop strings:
<|im_end|>
<|endoftext|>
<|end_of_text|>
<|end|>
<|im_start|>user
<|im_start|>system
Smoke-test sampler:
| field | value |
|---|---|
| temperature | 0.55 |
| top_p | 0.9 |
| top_k | 40 |
| min_p | 0.02 |
| repetition penalty | 1.08 |
| max new tokens | 192 |
Q8_0 GGUF:
huihui-qwen35-4b-roleplay-qlora-ms2048-s800-q8_0.gguf446723801C57DB417B255C90A068A39604400195DE1B97EE7B151A864434859BGGUF metadata readback:
| key | value |
|---|---|
general.architecture | qwen35 |
general.file_type | 7 |
general.quantization_version | 2 |
qwen35.block_count | 32 |
qwen35.context_length | 262144 |
qwen35.embedding_length | 2560 |
| tensor count | 426 |
| tensor types | 8:249, 0:177 |
The Q8_0 GGUF was loaded through the local EasyNovelAssistant
KoboldCpp.generate path on a temporary port with a safe Japanese writing
prompt. Checks passed:
<think> tagEvidence file:
eval/easynovel_s800_safe_prompt_eval.jsonfrom transformers import AutoModelForCausalLM, AutoTokenizer
repo = "zapabobouj/huihui-qwen35-4b-roleplay-qlora-ms2048-s800"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
repo,
trust_remote_code=True,
device_map="auto",
)
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = "huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated"
repo = "zapabobouj/huihui-qwen35-4b-roleplay-qlora-ms2048-s800"
tokenizer = AutoTokenizer.from_pretrained(f"{repo}/adapter")
model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, f"{repo}/adapter")
Use:
huihui-qwen35-4b-roleplay-qlora-ms2048-s800-q8_0.gguf
with the ChatML template above.
This is an adult-oriented Japanese roleplay fine-tune. It is not intended for
use with minors, non-consensual sexual content, or illegal activity. The local
training script applied an adult-consent lexical filter before SFT and
excluded the extra local JSONL source completely.
The smoke test verifies loadability and basic Japanese generation cleanliness; it is not a broad benchmark or a guarantee of generalization.