Stage-org/4b-diversity-LH-300-nyshot-27b-z-epoch3
Stage-org/4b-diversity-LH-300-nyshot-27b-z-epoch3 is a 4.5 billion parameter language model based on the Qwen/Qwen3.5-4B architecture, fine-tuned using a reinforcement learning (RL) method. This model was trained for 3 epochs on a diverse dataset, focusing on enhancing its capabilities through an advanced RL setup. It features a substantial context length of 32768 tokens, making it suitable for tasks requiring extensive contextual understanding and generation.
Loading preview...
Model Overview
Stage-org/4b-diversity-LH-300-nyshot-27b-z-epoch3 is a 4.5 billion parameter language model derived from the Qwen/Qwen3.5-4B base architecture. It has undergone reinforcement learning (RL) fine-tuning over 3 epochs, utilizing a specific dataset named Stage-org/4b-diversity-LH-300-nyshot-27b-z.
Key Training Details
- Base Model: Qwen/Qwen3.5-4B
- Training Method: Reinforcement Learning (RL)
- Epochs: 3
- Batch Size: 128
- Sequence Length: 300,000 (during training, though the model's context length is 32768 tokens)
- Optimization: AdamW optimizer with a learning rate of 1e-06 and Flash Attention 2 for efficiency.
- Generation Parameters: Features a temperature of 0.9 and a maximum token generation of 4096, with
enable_thinkingset to true, suggesting enhanced reasoning capabilities during generation. - Judge Model: Utilizes an external
gpt-5.6-lunamodel for open-ended judging during the RL process, indicating a focus on quality and nuanced evaluation of generated outputs.
Potential Use Cases
This model is particularly well-suited for applications that benefit from:
- Advanced Reasoning: The RL setup with an external judge and
enable_thinkingsuggests improved reasoning and response quality. - Long Context Processing: With a 32768 token context length, it can handle complex tasks requiring extensive input understanding.
- Generative Tasks: Its fine-tuning and generation parameters indicate a focus on producing high-quality, coherent, and contextually relevant text.