DCAgent/a1-toolscale
DCAgent/a1-toolscale is an 8 billion parameter language model fine-tuned from Qwen/Qwen3-8B, developed by DCAgent. It is specifically optimized for tool-use tasks, leveraging a specialized dataset for training. This model is designed to enhance performance in scenarios requiring complex reasoning and interaction with external tools, offering a 32768 token context length.
Loading preview...
Overview
DCAgent/a1-toolscale is an 8 billion parameter language model, fine-tuned from the Qwen/Qwen3-8B architecture. Developed by DCAgent, this model is specifically trained on the /e/scratch/jureap59/raoof1/sft_data/hf_hub/datasets--DCAgent--Toolscale-tasks-upsampled-10k_10k_glm_4.7_traces_jupiter/snapshots/6221a1d3f018d19e896374809ab80bfdecebd96f_thinking_preprocessed dataset. It features a substantial context length of 32768 tokens.
Key Training Details
The model underwent supervised fine-tuning with the following hyperparameters:
- Learning Rate: 4e-05
- Optimizer: ADAMW_TORCH_FUSED with betas=(0.9, 0.98) and epsilon=1e-08
- Epochs: 7.0
- Batch Size: A total training batch size of 16 (1 per device across 16 GPUs)
- Scheduler: Cosine learning rate scheduler with a 0.1 warmup ratio
Intended Use Cases
While specific intended uses and limitations are not detailed in the provided README, the training on a "Toolscale-tasks" dataset suggests its primary application is in scenarios requiring:
- Tool-use capabilities: Interacting with external APIs or functions.
- Complex reasoning: Tasks that benefit from structured thought processes, potentially involving multi-step problem-solving.
This model is likely best suited for applications where a robust understanding of instructions and the ability to generate tool-invoking code or structured outputs are critical.