yimn-Aghosh/zegrate-ai

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

The yimn-Aghosh/zegrate-ai model is a 14.8 billion parameter LoRA adapter checkpoint, fine-tuned from the Qwen2.5-14B-Instruct base model. This model utilizes QLoRA for efficient training, making it suitable for developers looking to integrate a specialized instruction-tuned variant of Qwen2.5-14B-Instruct into their applications. It offers enhanced performance for tasks aligned with its fine-tuning, providing a more tailored language generation capability.

Loading preview...

Zegrate 14B Phase 2 LoRA Checkpoint

This model is a LoRA (Low-Rank Adaptation) adapter checkpoint, representing Phase 2 of the Zegrate 14B training. It is built upon the robust Qwen2.5-14B-Instruct base model, offering a specialized instruction-tuned variant.

Key Characteristics & Training Details

  • Base Model: Qwen2.5-14B-Instruct
  • Parameter Count: 14.8 billion parameters
  • Context Length: 32,768 tokens
  • Training Method: QLoRA (Quantized LoRA), which combines 4-bit quantization with LoRA for efficient fine-tuning.
  • Training Steps: Fine-tuned over 400 steps.
  • Learning Rate: A learning rate of 2e-4 was used during training.
  • LoRA Rank: The LoRA adapter uses a rank of 32.

When to Use This Model

This LoRA checkpoint is ideal for developers who:

  • Are already using or considering the Qwen2.5-14B-Instruct base model and need a further specialized version.
  • Require a 14.8 billion parameter model with a large context window (32K tokens) for complex tasks.
  • Are looking for an efficiently fine-tuned model that leverages QLoRA for reduced memory footprint during adaptation.
  • Need a model that has undergone specific instruction-tuning, as indicated by its Phase 2 training, for improved performance on conversational or instruction-following tasks.