dusersad12/OrionLM-CheckpointRepo

OrionLM

---

1. Introduction

OrionLM is the newest release in our open language model series. For this release we scaled up the post-training compute and reworked the reinforcement-learning stage, which lifted OrionLM's reasoning and tool-use ability noticeably. The model was evaluated on a broad benchmark suite covering mathematics, code, and general logic, and it closes most of the gap to larger proprietary models.

Compared with the previous release, OrionLM handles multi-step reasoning far more reliably. On the GPQA Diamond set, pass@1 climbed from 41% in the previous version to 58% in this version. The gain lines up with longer thinking traces: the previous model averaged 9K tokens per problem on GPQA, while this version averages 18K tokens per problem.

Beyond raw reasoning, this release also tightens instruction-following and lowers the hallucination rate on factual prompts.

2. Evaluation Results

Comprehensive Benchmark Results

	Benchmark	Atlas-7B	Vega-1.3B	Atlas-7B-v2	OrionLM
Core Reasoning Tasks	Math Reasoning	0.462	0.488	0.495	0.506
	Logical Reasoning	0.711	0.725	0.740	0.731
	Common Sense	0.680	0.694	0.701	0.705
Language Understanding	Reading Comprehension	0.641	0.655	0.662	0.663
	Question Answering	0.555	0.571	0.578	0.584
	Text Classification	0.762	0.776	0.789	0.795
	Sentiment Analysis	0.739	0.755	0.768	0.772
Generation Tasks	Code Generation	0.581	0.595	0.604	0.600
	Creative Writing	0.531	0.548	0.556	0.557
	Dialogue Generation	0.592	0.605	0.612	0.611
	Summarization	0.712	0.728	0.735	0.739
Specialized Capabilities	Translation	0.761	0.776	0.784	0.788
	Knowledge Retrieval	0.621	0.638	0.648	0.653
	Instruction Following	0.701	0.718	0.725	0.730
	Safety Evaluation	0.688	0.704	0.712	0.717

Overall Performance Summary

OrionLM shows consistent gains across every benchmark category, with the largest jumps in reasoning and generation tasks.

3. Chat Website & API Platform

We host a chat interface and an API endpoint so you can try OrionLM directly. See our official website for details.

4. How to Run Locally

Refer to our code repository for full instructions on running OrionLM locally.

A few usage notes compared with earlier releases:

1. A system prompt is now supported.
2. You no longer need to prepend a special token to force a thinking mode.

OrionLM-Small shares the architecture and tokenizer of the base OrionLM model and can be run the same way as its base model.

System Prompt

We recommend pairing OrionLM with a dated system prompt.

You are OrionLM, a helpful AI assistant.
Today is {current date}.

For example,

You are OrionLM, a helpful AI assistant.
Today is Sep 27, 2026, Sunday.

Temperature

We recommend setting the temperature parameter $T_{model}$ to 0.7.

Prompts for File Uploading and Web Search

For file uploading, use the template below where {file_name}, {file_content} and {question} are arguments.

file_template = \
"""[file name]: {file_name}
[file content begin]
{file_content}
[file content end]
{question}"""

For web-search-augmented generation, use the template below where {search_results}, {cur_date}, and {question} are arguments.

search_answer_en_template = \
'''# The following contents are the search results related to the user's message:
{search_results}
In the search results I provide to you, each result is formatted as [webpage X begin]...[webpage X end], where X represents the numerical index of each article. Please cite the context at the end of the relevant sentence when appropriate. Use the citation format [citation:X] in the corresponding part of your answer. If a sentence is derived from multiple contexts, list all relevant citation numbers, such as [citation:3][citation:5]. Be sure not to cluster all citations at the end; instead, include them in the corresponding parts of the answer.
When responding, please keep the following points in mind:
- Today is {cur_date}.
- Not all content in the search results is closely related to the user's question. You need to evaluate and filter the search results based on the question.
- For listing-type questions (e.g., listing all flight information), try to limit the answer to 10 key points and inform the user that they can refer to the search sources for complete information. Prioritize providing the most complete and relevant items in the list. Avoid mentioning content not provided in the search results unless necessary.
- For creative tasks (e.g., writing an essay), ensure that references are cited within the body of the text, such as [citation:3][citation:5], rather than only at the end of the text. You need to interpret and summarize the user's requirements, choose an appropriate format, fully utilize the search results, extract key information, and generate an answer that is insightful, creative, and professional. Extend the length of your response as much as possible, addressing each point in detail and from multiple perspectives, ensuring the content is rich and thorough.
- If the response is lengthy, structure it well and summarize it in paragraphs. If a point-by-point format is needed, try to limit it to the 5 points and merge related content.
- For objective Q&A, if the answer is very brief, you may add one or two related sentences to enrich the content.
- Choose an appropriate and visually appealing format for your response based on the user's requirements and the content of the answer, ensuring strong readability.
- Your answer should synthesize information from multiple relevant webpages and avoid repeatedly citing the same webpage.
- Unless the user requests otherwise, your response should be in the same language as the user's question.
# The user's message is:
{question}'''

5. License

This code repository is licensed under the Apache-2.0 License. Use of the OrionLM models is also governed by the Apache-2.0 License. The model series supports commercial use and distillation.

6. Contact

If you have any questions, please open an issue on our GitHub repository or reach us at contact@orionlm.ai.