yuzhounie/sft_qwen32b

TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Jan 4, 2026License:otherArchitecture:Transformer Cold

The yuzhounie/sft_qwen32b model is a 32.8 billion parameter language model, fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct. It was trained on the tb3000_agent_diverse_real dataset over 5 epochs. This model is specifically optimized for agent-based tasks and real-world diverse scenarios, leveraging its large parameter count for complex instruction following.

Loading preview...

Model Overview

The yuzhounie/sft_qwen32b model is a 32.8 billion parameter language model, derived from the Qwen/Qwen2.5-Coder-32B-Instruct base model. It has been specifically fine-tuned on the tb3000_agent_diverse_real dataset, indicating an optimization for agent-based applications and handling diverse, real-world instructions.

Training Details

The model underwent 5 epochs of training with a learning rate of 5e-06. Key hyperparameters included a train_batch_size of 1, gradient_accumulation_steps of 4, resulting in a total_train_batch_size of 32. The optimizer used was adamw_torch_fused with cosine learning rate scheduling and a warmup ratio of 0.1. This training regimen suggests a focus on robust performance across varied tasks.

Potential Use Cases

Given its fine-tuning on an "agent-diverse-real" dataset, this model is likely well-suited for:

  • Complex Agentic Workflows: Handling multi-step instructions and decision-making in automated agents.
  • Diverse Instruction Following: Interpreting and executing a wide range of user prompts in real-world contexts.
  • Code-Related Agent Tasks: Leveraging its Coder base for agentic tasks that involve code generation, understanding, or execution.