uniform484_baseline_global_step_847

uniform484_baseline_global_step_847

Checkpoint exported in Hugging Face format (bfloat16, Qwen3ForCausalLM).

Load with transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "mskm3266/uniform484_baseline_global_step_847"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "Solve: what is the remainder when 2^100 is divided by 7?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Load with vLLM

from vllm import LLM, SamplingParams

llm = LLM(model="mskm3266/uniform484_baseline_global_step_847", dtype="bfloat16")
out = llm.generate(
    ["Solve: what is the remainder when 2^100 is divided by 7?"],
    SamplingParams(temperature=0.7, top_p=0.95, max_tokens=512),
)
print(out[0].outputs[0].text)

Or serve it:

vllm serve mskm3266/uniform484_baseline_global_step_847 --dtype bfloat16

Download the files only

hf download mskm3266/uniform484_baseline_global_step_847 --local-dir ./uniform484_baseline_global_step_847

Notes

• This is a base-style model: it has no chat template applied by default, so feed it raw prompts (the tokenizer does ship a chat_template.jinja inherited from the base model if you want it).
• Weights are stored in bfloat16; loading in float16 is not recommended.