This model is a fine-tuned version of
Qwen3-4B using
ReMax without KL penalty for mathematical reasoning.
ReMax uses a greedy-decoded response's reward as the baseline for advantages.
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model = AutoModelForCausalLM.from_pretrained("jaygala24/Qwen3-4B-ReMax-math-reasoning", revision="step-0200") # optional branch, e.g. "step-0400"
4tokenizer = AutoTokenizer.from_pretrained("jaygala24/Qwen3-4B-ReMax-math-reasoning", revision="step-0200")
5
6prompt = "Please reason step by step, and put your final answer within \\boxed{}.\n\nWhat is the sum of 123 and 456?"
7inputs = tokenizer(prompt, return_tensors="pt")
8outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.7)
9print(tokenizer.decode(outputs[0], skip_special_tokens=True))
1from vllm import LLM, SamplingParams
2
3llm = LLM(model="jaygala24/Qwen3-4B-ReMax-math-reasoning", revision="step-0200") # optional branch, e.g. "step-0400"
4sampling_params = SamplingParams(temperature=0.7, max_tokens=4096)
5
6prompt = "Please reason step by step, and put your final answer within \boxed{}.
7
8What is the sum of 123 and 456?"
9outputs = llm.generate([prompt], sampling_params)
10print(outputs[0].outputs[0].text)
