Qwen2.5-0.5B-HiddenDistilled

This repository contains the fully merged base + adapter weights for Qwen2.5-0.5B-Instruct distilled from the teacher model Qwen2.5-3B-Instruct. The distillation pipeline optimizes a composite objective combining supervised learning cross-entropy, logit-level KL divergence, and MSE alignment of projected hidden states.

• Base Student Model: Qwen/Qwen2.5-0.5B-Instruct
• Teacher Model: Qwen/Qwen2.5-3B-Instruct
• LoRA Adapter Repo: sarimahsan101/Qwen2.5-0.5B-HiddenDistilled-LoRA
• Training Code (GitHub): sarimahsan101/distillation-hiddenstates

📊 Trial Run Evaluation Metrics

Evaluation metrics compiled on a single NVIDIA Tesla T4 (16GB) GPU for 1 epoch on a subset of Dolly, Alpaca SFT, and Ultrachat:

Metric	Before Distillation	After Distillation	Change	Status
Validation Perplexity	5.0924	5.2620	+0.1696	✗
Teacher-Student KL Divergence	2.7913	1.9637	-0.8276	✓
Hidden State Cosine Similarity	0.0075	0.0054	-0.0021	✗

🚀 How to Use (Merged Model)

You can load this model directly using standard Hugging Face Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "sarimahsan101/Qwen2.5-0.5B-HiddenDistilled"

# Load tokenizer and merged model
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
model.eval()

# Inference example
messages = [{"role": "user", "content": "Explain gravity in one sentence."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=50)

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

🛠️ Training Configurations & Details

• Framework: PyTorch & Hugging Face Transformers / Trainer
• Quantization: 4-bit NF4 double quantization (bitsandbytes) for training, adapter weights were merged with FP16 base student.
• Loss Weights: Cross-Entropy: 0.3, KL Divergence: 0.4, Hidden MSE: 0.3