You can deploy this model and start earning money today!

Phonsiri/
Qwen3.5-0.8B-Base-Distillation-Qwen3.5-9B

knowledge-distillation

reasoning

Phonsiri/Qwen3.5-Distillation-Dataset

Qwen/Qwen3.5-0.8B-Base

finetune

apache-2.0

Views

No views yet

Qwen3.5 0.8B (Phase 1: Supervised Fine-Tuning)

This model serves as the Phase 1 Supervised Fine-Tuning (SFT) checkpoint in a Two-Phase Knowledge Distillation pipeline. It is built upon Qwen/Qwen3.5-0.8B-Base and fine-tuned to learn the structural formatting and reasoning chains (<think> tags) generated by a larger teacher model (Qwen/Qwen3.5-9B).

🚨 Note: This is an intermediate model. For the final, fully-distilled model with aligned probabilities (Logits Distillation), please use Phonsiri/Qwen3.5-0.8B-Distillation-Phase2.

🚀 Model Details

Base Model: Qwen/Qwen3.5-0.8B-Base
Teacher Configuration: Qwen/Qwen3.5-9B (Used for dataset generation)
Language(s): English (Primary), Thai
Architecture: Causal Language Modeling (Decoder-only)
License: Apache 2.0

🧪 Training Methodology

This model was trained exclusively offline (Off-Policy) using standard Cross Entropy (CE) Loss.

A custom dataset of 7,500 mixed-domain prompts (comprising Math, General Instructions, and Coding tasks) was fed into the 9B Teacher model to generate high-quality ground-truth responses. The generation explicitly enabled "Thinking Mode" to force the teacher to output step-by-step reasoning encapsulated in <think> ... </think> tags before answering.

The 0.8B-Base student model was then fine-tuned on this dataset to adapt to the instruction format and mimic the presence of reasoning chains. This warm-up phase is critical to ensure both the Teacher and Student output formats are perfectly synchronized before the intensive On-Policy KL-Divergence Distillation in Phase 2.

📊 Hyperparameters (SFT)

Epochs: 2 (Intentional early stopping to prevent severe overfitting)
Optimizer: AdamW
Learning Rate: 2e-5 (with Cosine scheduling)
Batch Size (Effective): 32
Precision: fp16
Max Sequence Length: 7000

💻 How to Use (Inference)

You can test this intermediate model to observe how it structures its reasoning process.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Phonsiri/Qwen3.5-0.8B-Base-Distillation-Qwen3.5-9B"
subfolder = "step_100" # Use the desired checkpoint step

tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    subfolder=subfolder,
    torch_dtype=torch.float16, 
    device_map="auto"
)

# Example Prompt
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is the square root of 256? Please explain your thinking."}
]

text = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True,
    enable_thinking=True, # 💡 Command the model to use <think> reasoning
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))