CromonHarry/qwen3-32b-creativity

qwen3-32b Fine-tuned for Story Creativity Evaluation

Fine-tuned qwen3-32b model for evaluating story creativity (1-5 scale).

Model Details

• Base Model: Qwen/qwen3-32b
• Method: LoRA (r=8, alpha=16)
• Training Data: 5000 story creativity evaluations
• Epochs: 3
• Final Loss: ~1.26

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# Load models
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/qwen3-32b",
    torch_dtype=torch.float16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "CromonHarry/qwen3-32b-creativity")
tokenizer = AutoTokenizer.from_pretrained("CromonHarry/qwen3-32b-creativity")

# Evaluate a story
prompt = '''Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Evaluate the creativity of the following story. You need to output the following format:
[Your review]
Creativity score: [A float score from 1 to 5]

### Input:
{your_story_here}

### Response:
'''

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Configuration

• Learning Rate: 2e-4 (cosine scheduler)
• Batch Size: 4 (effective: 32)
• Optimizer: AdamW
• Gradient Checkpointing: Enabled