Qwen2.5-0.5B-HiddenDistilled-LoRA

This model repository contains the LoRA adapter weights for Qwen2.5-0.5B-Instruct distilled from the teacher model Qwen2.5-3B-Instruct. The distillation objective integrates SFT cross-entropy loss, logit-level Kullback-Leibler (KL) divergence, and linear projection-based hidden state matching.

• Base Student Model: Qwen/Qwen2.5-0.5B-Instruct
• Teacher Model: Qwen/Qwen2.5-3B-Instruct
• Merged Distilled Model: sarimahsan/distillation-hiddenstates-code

📊 Trial Run Evaluation Metrics

These metrics were recorded during a trial run on a single NVIDIA Tesla T4 (16GB) GPU for 1 epoch on a subset of Alpaca, Dolly, 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	✗

Analysis:

• The Teacher-Student KL Divergence improved by 0.8276, showing that the student's output probabilities are starting to align with the teacher's logits.
• Perplexity and Hidden similarity did not improve during the brief training run due to the restricted dataset size and quick epoch limits. A larger compute and training duration are recommended to achieve full convergence of the projection layers.

🚀 How to Use (LoRA Adapter)

To load and use this adapter, run:

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

base_model_name = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_name = "sarimahsan101/Qwen2.5-0.5B-HiddenDistilled-LoRA"

# Load tokenizer and base model
tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, adapter_name)
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)
• PEFT Method: LoRA (Rank = 32, Alpha = 64, Target modules = all linear layers)
• Optimization: Paged AdamW 8bit with Cosine LR scheduler
• Loss Weights: Cross-Entropy: 0.3, KL Divergence: 0.4, Hidden MSE: 0.3