First, supervised fine-tuning (SFT) adapts the base checkpoint on HuggingFaceH4/ultrachat_200k using the train_sft split. That stage used NVIDIA H100 NVL hardware with 1,024-token sequences, batch size 24, gradient accumulation 2, learning rate 2e-5, and 100 warmup steps.

Second, direct preference optimization (DPO) refines the SFT checkpoint on KatoHF/chatbot_arena_binarized with the chat_refine_strict recipe. That stage used NVIDIA H100 PCIe hardware with 1,024-token sequences, batch size 4, gradient accumulation 8, learning rate 5e-6, beta 0.2, and 10 warmup steps.

The published checkpoint is stored in bfloat16 and split across 5 safetensor shards for easier loading.

Training data

Base checkpoint: Argonne 2.5-base
SFT data: HuggingFaceH4/ultrachat_200k (train_sft)
DPO data: KatoHF/chatbot_arena_binarized (chat_refine_strict)

Tokenizer

This model uses the Qwen3 tokenizer family via the Qwen2Tokenizer compatibility class.

Source code

The release was built from the GitHub main branch codebase: https://github.com/PursuitOfDataScience/ArgonneAI/tree/main

Key scripts:

Recommended inference config

Item	Value
Context length	1,024 tokens
Temperature	0.8
Top-p	0.9
Repetition penalty	1.3
No-repeat n-gram size	4
Seed	444

These settings are the recommended defaults for inference.

Inference

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "PursuitOfDataScience/Argonne2.5-instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)

prompt = "Write a short paragraph about scientific computing at Argonne National Laboratory."
inputs = tokenizer(prompt, return_tensors="pt")
input_ids = inputs["input_ids"].to(device)

seed = 444
torch.manual_seed(seed)
if device.startswith("cuda"):
    torch.cuda.manual_seed_all(seed)

output_ids = model.generate(
    input_ids,
    max_length=input_ids.shape[1] + 128,
    temperature=0.8,
    top_p=0.9,
    do_sample=True,
    repetition_penalty=1.3,
    no_repeat_ngram_size=4,
)
print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Usage notes

Load with trust_remote_code=True.
The custom generate method accepts repetition_penalty and no_repeat_ngram_size.
The sweep-derived repetition controls are available in the repository's custom generation loop, not the checkpoint's built-in generate method.
Weights are published as 5 bf16 safetensor shards.
The instruct checkpoint inherits the base tokenizer and chat template.

Citation

bibtex

@misc{argonne25instruct,
  author = {PursuitOfDataScience},
  title = {Argonne 2.5-instruct},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/PursuitOfDataScience/Argonne2.5-instruct}
}