Amu/t1-1.5B

It's a 1.5B model.

It's a distill model like s1 and deepseek-r1-distill.

It's test model. I hope I can reproduce a rl model like RL-Zero.

This model is a mini-step.

Thanks for evveryone in the open community.

how to use:

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model = LLM(
    "Amu/t1-1.5B"
)
tok = AutoTokenizer.from_pretrained("simplescaling/s1-32B")

stop_token_ids = tok("<|im_end|>")["input_ids"]

sampling_params = SamplingParams(
    max_tokens=32768,
    min_tokens=0,
    stop_token_ids=stop_token_ids,
)

prompt = "How many r in raspberry"
prompt = "<|im_start|>system\nYou are t1, created by Amu. You are a helpful assistant.<|im_end|>\n<|im_start|>user\n" + prompt + "<|im_end|>\n<|im_start|>assistant\n"

o = model.generate(prompt, sampling_params=sampling_params)
print(o[0].outputs[0].text)