dusersad12/HelixLM-PreviewRepo

HelixLM

---

1. Introduction

HelixLM is the first release in the Helix series of open language models trained by the Helix Lab. The model was post-trained with an interleaved reasoning pipeline and shows particular strength in mathematics, code understanding, and multi-step logical problems. Across our internal suites it lands competitively with several much larger baselines.

Compared with the previous internal candidate, HelixLM dedicates substantially more thinking budget to hard problems before committing to an answer. On a held-out reasoning mix, average token usage per problem roughly doubled while answer quality improved by a clear margin, which matches the trend we hoped for when scaling up deliberate reasoning.

The release candidate also ships with a more conservative safety tuning pass and a refreshed function-calling head.

2. Evaluation Results

Comprehensive Benchmark Results

	Benchmark	BaselLM	Prism-7B	Prism-13B	HelixLM
Core Reasoning Tasks	Math Reasoning	0.482	0.507	0.514	0.537
	Logical Reasoning	0.771	0.784	0.793	0.801
	Common Sense	0.708	0.694	0.717	0.727
Language Understanding	Reading Comprehension	0.659	0.673	0.678	0.689
	Question Answering	0.564	0.581	0.587	0.600
	Text Classification	0.796	0.804	0.813	0.820
	Sentiment Analysis	0.762	0.770	0.779	0.786
Generation Tasks	Code Generation	0.601	0.617	0.626	0.636
	Creative Writing	0.571	0.562	0.584	0.595
	Dialogue Generation	0.608	0.622	0.630	0.634
	Summarization	0.729	0.739	0.744	0.759
Specialized Capabilities	Translation	0.755	0.772	0.778	0.800
	Knowledge Retrieval	0.622	0.639	0.641	0.670
	Instruction Following	0.711	0.727	0.731	0.750
	Safety Evaluation	0.692	0.675	0.699	0.732

Overall Performance Summary

HelixLM holds up well in every evaluated category, with the clearest headroom shown on the reasoning-heavy splits.

3. Chat Website & API Platform

We host a playground and an API endpoint so anyone can try HelixLM directly. Details are on the Helix Lab website.

4. How to Run Locally

Please refer to our code repository for more information about running HelixLM locally.

Compared with previous internal candidates, the usage recommendations for HelixLM are:

1. System prompt is supported.
2. Special tokens are no longer needed to force the model into a specific thinking pattern.

The architecture of HelixLM-Small matches its base model, and it shares the same tokenizer configuration as the main HelixLM release. It can be loaded the same way as its base model.

System Prompt

We recommend using the following system prompt with a specific date.

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

For example,

You are HelixLM, a helpful AI assistant.
Today is May 28, 2025, Monday.

Temperature

We recommend setting the temperature parameter $T_{model}$ to 0.6.

Prompts for File Uploading and Web Search

For file uploading, please follow the template to create prompts, 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 enhanced generation, we recommend the following prompt template 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 the answer to 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. The use of HelixLM models is also subject to the Apache 2.0 License. The model series supports commercial use and distillation.

6. Contact

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