SuperAPI AI Directory
Partner with us中文AI newsFavorites

AI developer platforms

3,641 results · Page 94 of 102

simple_GRPO

AI developer platforms

simple_GRPO is a lightweight open-source GRPO implementation for reproducing the reasoning capabilities of R1-like large language models. The code is only about 200 lines, spread across two files, and depends on standard libraries such as deepspeed and torch, with no need for ray. It draws on the core loss calculation formula from Hugging Face trl, and supports saving GPU memory, quickly understanding the RL process, trying multi-answer generation, regrouping, KL penalty, and parameter tuning, among other things. An "Aha moment" can be observed in the early stages of training. The project provides a decoupled reference model design, allowing it to run on different GPUs or even different machines, and supports vLLM-accelerated inference and the reinforce++ algorithm. Experiments show that training is completed within 1 hour on an A800, and the Qwen2.5-7B and 3B models exhibit an Aha moment within 30 steps, with accuracy stably above 60% and format compliance reaching 100%.

Read more

simpleRL-reason

AI developer platforms

simpleRL-reason is an open-source reinforcement learning training codebase for enhancing reasoning ability, released on GitHub by the NLP team at the Hong Kong University of Science and Technology. It provides a simple reinforcement learning training scheme that can improve the reasoning ability of base models using only rule-based rewards and the GSM8K/Math datasets. The project has successfully conducted zero RL training on 10 different base models, including Llama3 8B, Mistral 7B/24B, DeepSeekMath 7B, Qwen2.5 0.5B/1.5B/7B/14B/32B, and Qwen2.5-Math-7B. Using limited data of 8K examples, the accuracy improvement reaches 10 to more than 20 absolute percentage points. The project has open-sourced the code, model checkpoints, and analysis tools, and provides related papers and blogs. The code is implemented based on Verl, supports custom environment settings and FSDP training, uses the GRPO algorithm, and accelerates training with Ray and vLLM.

Read more