Neural-Hacker/Qwen3.5-2B-chat

Qwen3.5-2B Chat LoRA (Ultrachat SFT)

Overview

• LoRA fine-tuned model based on Qwen/Qwen3.5-2B
• Trained for chat-style instruction following
• Uses conversational formatting (<|user|>, <|assistant|>)
• Optimized for fast training with a reduced dataset subset

Important

• This repository contains LoRA adapter weights only
• Base model is NOT included
• Required base model: Qwen/Qwen3.5-2B

Usage

• Load base model and adapter
• Use chat template for proper responses

from transformers import AutoTokenizer, AutoModelForCausalLM  
from peft import PeftModel  

base_model = "Qwen/Qwen3.5-2B"  
adapter = "your-username/your-repo-name"  

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)  

model = AutoModelForCausalLM.from_pretrained(  
    base_model,  
    torch_dtype=torch.float16,  
    device_map="auto",  
    trust_remote_code=True  
)  

model = PeftModel.from_pretrained(model, adapter)  
model.eval()  

messages = [  
    {"role": "user", "content": "Explain gravity briefly"}  
]  

text = tokenizer.apply_chat_template(  
    messages,  
    tokenize=False,  
    add_generation_prompt=True  
)  

inputs = tokenizer(text, return_tensors="pt").to(model.device)  

outputs = model.generate(  
    **inputs,  
    max_new_tokens=400,  
    do_sample=True,  
    temperature=0.7,  
    top_p=0.9,  
    eos_token_id=tokenizer.eos_token_id  
)  

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

if "</think>" in response:  
    response = response.split("</think>")[-1]  

response = response.replace(text, "").strip()  

print(response)

Training Details

• Base Model: Qwen/Qwen3.5-2B
• Method: LoRA (PEFT)
• Framework: Transformers + TRL (SFTTrainer)
• Dataset: HuggingFaceH4/ultrachat_200k (train_sft split)
• Data Used: 10,000 samples (subset for fast training)
• Task: Chat-style supervised fine-tuning

Hyperparameters

• Max Steps: 1000
• Max Sequence Length: 1024
• Batch Size: 4
• Gradient Accumulation: 4
• Learning Rate: 2e-4
• Precision: bfloat16
• Gradient Checkpointing: Enabled

Capabilities

• Multi-turn conversational responses
• Instruction-following chat behavior
• Handles both short and long-form answers
• Supports English and Hindi inputs (inherited from base model)

Limitations

• Trained on a subset (10k samples), not full dataset
• May produce inconsistent or incomplete responses
• Not fully aligned for safety or factual accuracy
• May still show verbosity or repetition

Example

• Input: What is gravity?
• Output: Gravity is a fundamental force that attracts objects with mass toward each other.

Model Type

• Architecture: Causal Language Model
• Task: Text Generation (Chat-style)
• Fine-tuning: LoRA Adapter

License

This model is released under the CC BY-NC 4.0 license.