Model Summary
Atla Selene 1 is a state-of-the-art large language model-as-a-judge. Selene 1 achieves frontier-level performance across 11 evaluation benchmarks, outperforming OpenAI's o1, o3-mini, and GPT-4o, Anthropic's Claude 3.5 Sonnet, Meta's Llama 3.3, and DeepSeek's R1.
Selene 1 was post-trained from Llama-3.3-70B on a combined SFT+DPO objective using the same recipe as Selene Mini. Trained across a wide range of evaluation tasks and scoring criteria, Selene 1 outperforms frontier models across 11 benchmarks covering three different types of tasks:
- Absolute scoring, e.g. "Evaluate the harmlessness of this response on a scale of 1-5"
- Classification, e.g. "Does this response address the user query? Answer Yes or No."
- Pairwise preference. e.g. "Which of the following responses is more logically consistent - A or B?"
Selene 1 particularly excels at capturing human judgements on nuanced, complex real-world evaluations, as evidenced by its state-of-the-art performance on FLASK, MT-Bench, RewardBench and Auto-J.
If you're interested in evaluating more complex systems like LLM agents, learn more here.
Model Details
- Developed by: Atla
- Model type: Post-trained from Llama-3.3-70B
- Parameters: 70 billion
- Language(s) (NLP): Primarily English but supports German, French, Italian, Portuguese, Hindi, Spanish, Thai
- Context length 128K
Model Use
Selene 1 can be used as a general-purpose evaluation model. It supports different inputs & scoring scales, generates structured evaluation outputs, and provides qualitative critiques with reasoning. The model is highly steerable with customizable evaluation criteria and can evaluate responses with or without reference responses.
Try our cookbooks to get started with two popular use cases below:
To achieve best results, we provide the prompts we used for training here.
Remember to apply the conversation template of Llama 3 - not doing so might lead to unexpected behaviors. You can find the conversation class at this link or you can refer to the below code that will apply it.
Quickstart (HF Transformers):
from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto
model_id = "AtlaAI/Selene-1-Llama-3.3-70B"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)
prompt = "I heard you can evaluate my responses?" # replace with your prompt / we provide prompt templates used during training at github.com/atla-ai/selene-mini/tree/main/prompt-templates
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512, do_sample=True)
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
Contact
support@atla-ai.com You can also join our Discord!
Citation
If you are using the model, cite using
@misc{alexandru2025atlaseleneminigeneral,
title={Atla Selene Mini: A General Purpose Evaluation Model},
author={Andrei Alexandru and Antonia Calvi and Henry Broomfield and Jackson Golden and Kyle Dai and Mathias Leys and Maurice Burger and Max Bartolo and Roman Engeler and Sashank Pisupati and Toby Drane and Young Sun Park},
year={2025},
eprint={2501.17195},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2501.17195},
}