Modelpublic

tvist1988/happynable-8b-nsfw-en-v0.2-full

sourceHugging Faceupdated 1y agoView on Hugging Face

0likes7downloads

Model Card

📒 Model Card: tvist1988/happynable-8b-nsfw-en-v0.2-full

Important! This model was created as part of fine-tuning training. I would greatly appreciate your feedback..

Model Description

tvist1988/happynable-8b-nsfw-en-v0.2-full is an experimental fine-tuned 8‑billion parameter language model optimized for balanced story writing, roleplaying. It handles both SFW and NSFW content but may produce unexpected or incorrect outputs due to the nature of fine‑tuning.

—Base architecture: Qwen3 8b
—Version: v0.2-full
—Use cases:

—Creative story generation (SFW & NSFW)
—Multi‑turn roleplay

Training Details

—Compute: 4× NVIDIA RTX4090, \~9 hours total over multiple runs
—Method: QLORA
—Data sources:

—Mixed SFW & NSFW story‑writing datasets (contributions by Gryphe)

Wandb

Disclaimer

This model is provided as an experimental fine‑tuning of a base model. It may not always behave as expected or produce correct responses. Use with caution and verify outputs before relying on them.

Recommended Sampling Settings

Temperature: 1.12 – 1.5  
Min‑P: 0.01  
Top‑K: 40 
Repetition Penalty: 1.5

Stopping Criteria

Prompt Template

<|im_start|>system
{system_prompt}
<|im_end|>

<|im_start|>user
{input}
<|im_end|>

<|im_start|>assistant
{output}
<|im_end|>

Axolotl Configuration

yaml

base_model: Qwen/Qwen3-8B

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
strict: false
load_in_8bit: false
load_in_4bit: true

chat_template: chatml
datasets:
  - path: Gryphe/Opus-WritingPrompts
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
  - path: Gryphe/Sonnet3.5-Charcard-Roleplay
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
shuffle_merged_datasets: true

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/lora-out

sequence_len: 8192
sample_packing: false
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.1
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - down_proj
  - up_proj

lora_target_linear: true
peft_use_rslora: true

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 5e-5
auto_batch_size: true
train_on_inputs: false
bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true
auto_resume_from_checkpoints: true
warmup_ratio: 0.05
evals_per_epoch: 4
saves_per_epoch: 4
weight_decay: 0.01

save_total_limit: 3

deepspeed: deepspeed_configs/zero1.json

Feedback & Support & Support

Thank you for trying tvist1988/happynable-8b-nsfw-en-v0.2-full! Your feedback is invaluable.