DCAgent/g1_clean_hybrid_25k_32b
DCAgent/g1_clean_hybrid_25k_32b is a 32 billion parameter language model fine-tuned from Qwen/Qwen3-32B. This model is specifically adapted using a unique dataset focused on 'thinking preprocessed' traces, suggesting an optimization for complex reasoning or agentic workflows. It is designed for applications requiring advanced understanding and generation capabilities derived from its specialized training data.
Loading preview...
Model Overview
DCAgent/g1_clean_hybrid_25k_32b is a 32 billion parameter language model, fine-tuned from the robust Qwen/Qwen3-32B architecture. Its development focused on a specialized dataset, /e/scratch/jureap59/raoof1/sft_data/hf_hub/datasets--DCAgent--g1_clean_hybrid_scaffold_25k_glm47_traces/snapshots/ad622359a4cfbac08ec8e7bbe09f4f41a72a1834_thinking_preprocessed, indicating a targeted approach to enhance specific cognitive or processing capabilities.
Training Details
The model underwent a supervised fine-tuning (SFT) process with the following key hyperparameters:
- Learning Rate: 4e-05
- Optimizer: AdamW (betas=(0.9, 0.98), epsilon=1e-08)
- Batch Size: 1 (per device), totaling 96 across 96 devices
- Epochs: 5.0
- Scheduler: Cosine with 0.1 warmup ratio
Potential Use Cases
Given its fine-tuning on a 'thinking preprocessed' dataset, this model is likely optimized for:
- Complex Reasoning Tasks: Applications requiring multi-step thought processes or logical deduction.
- Agentic Workflows: Scenarios where the model needs to simulate internal thought or planning.
- Specialized Data Processing: Tasks aligned with the unique characteristics of its training data, potentially involving structured or semi-structured information analysis.