📌 Model ID: xsanskarx/thinkygemma-4b
📌 Parameters trained: 1.8 billion
📌 Trained on: 25k rows of verified Chain-of-Thought (CoT) traces from DeepSeek R1 and Qwen QWQ
📌 Next planned step: GRPO 📌 adapters repo: xsanskarx/thinkgemma-4b

Model Description

This is a fine-tuned version of Google's Gemma-3-4b-it, adapted for **structured reasoning / induced reasoning behaviour.

Training Details

Hardware: Single NVIDIA H100
Training Time: 9 hours
Training Method: LoRA fine-tuning (r = 128, alpha = 256)
Dataset: 25k CoT traces
Base Model: google/gemma-3-4b-it

Setup

python

from transformers import AutoTokenizer, Gemma3ForConditionalGeneration, TextStreamer
import torch

# Load model and tokenizer
model_id = "xsanskarx/thinkygemma-4b"
model = Gemma3ForConditionalGeneration.from_pretrained(model_id, device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(model_id)

def ask_model(prompt: str, max_tokens=8192, temperature=0.7):
    """
    Function to ask a question to the model and stream the response.
    """
    messages = [
        {"role": "system", "content": "You are an expert math problem solver, think and reason inside <think> tags, enclose all reasoning in <think> tags, verifying logic step by step and then return your final structured answer"},
        {"role": "user", "content": prompt}
    ]

    formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False)
    inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)

    streamer = TextStreamer(tokenizer, skip_special_tokens=True)
    with torch.inference_mode():
        model.generate(**inputs, max_new_tokens=max_tokens, do_sample=True, temperature=temperature, streamer=streamer)

# Example usage
ask_model("do 2+2")