SuperAPI AI Directory
Partner with us中文AI newsFavorites

AI developer platforms

3,641 results · Page 84 of 102

Open-O1

AI developer platforms

Open-O1 is an open-source project aimed at reproducing the long reasoning and problem-solving capabilities of the OpenAI O1 model through open-source innovation. The project compiles a set of SFT data for chain-of-thought (CoT) activation and uses it to train LLaMA and Qwen models, enabling smaller models to achieve enhanced long reasoning performance. The project provides two models, OpenO1-LLaMA-8B-v0.1 and OpenO1-Qwen-7B-v0.1, as well as the OpenO1 SFT data for CoT activation, all available on HuggingFace. Official evaluations show that its models achieve competitive results on benchmarks such as GSM8K, MATH, MMLU, ARC-C, and BBH, with some metrics outperforming llama3.1-8b-instruct. The project is non-profit, welcomes community participation, and will continue to open-source the code and materials from the development process.

Read more

open-r1-multimodal

AI developer platforms

open-r1-multimodal is a fork of the Hugging Face open-r1 project, aiming to extend the DeepSeek-R1 paradigm to multimodal models and add support for multimodal model training. The project implements multimodal R1 based on huggingface/open-r1 and deepseek-ai/DeepSeek-R1, integrates vision-language models (VLMs) available in transformers such as the Qwen2-VL series and Aria-MoE, and uses the GRPO algorithm for training. The project has open-sourced the first batch of 8k multimodal RL training examples, focusing on mathematical reasoning, with data generated by GPT4o based on Math360K and Geo170K, including reasoning paths and verifiable answers; it has also open-sourced the lmms-lab/Qwen2-VL-2B-GRPO-8k and lmms-lab/Qwen2-VL-7B-GRPO-8k models trained with GRPO. The repository provides scripts for data inspection and custom generation, as well as reference training commands for an 8x H100 GPU node.

Read more

open-r1-reprod

AI developer platforms

open-r1-reprod is a reproduction project of open-r1, using GRPO to train Qwen models of 0.5B, 1.5B, 3B, and 7B on the MATH-lighteval math dataset, observing and analyzing phenomena such as the reasoning abilities, rule reward hacks, and multilingual mixed reasoning of models of different sizes. The project code is based on open-r1, modifying the config_grpo.yaml configuration file, training uses a cosine learning rate (maximum 3.0e-06), and the rewards include accuracy reward, format reward, and tag_count_reward. The project is not exactly a success, but there are some interesting findings, for example, the 0.5B small model is very difficult to learn reasoning ability through GRPO, models of 1.5B and above begin to take effect; the 7B model's loss curve fluctuates violently, and phenomena such as garbled output appear.

Read more