ermiaazarkhalili/FastContext-4B-RL_base-Function-Calling-xLAM-Unsloth

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 21, 2026Architecture:Transformer Featherless Exclusive Cold

FastContext-4B-RL_base-Function-Calling-xLAM-Unsloth is a 4.0 billion parameter Qwen3ForCausalLM architecture model, fine-tuned by ermiaazarkhalili. It is a LoRA fine-tune of microsoft/FastContext-1.0-4B-RL, specifically optimized for function calling tasks. The model was supervised fine-tuned using the Salesforce/xlam-function-calling-60k dataset, making it suitable for applications requiring structured tool use and API interaction.

Loading preview...

Model Overview

This model, FastContext-4B-RL_base-Function-Calling-xLAM-Unsloth, is a 4.0 billion parameter language model based on the Qwen3ForCausalLM architecture. It is a LoRA fine-tune of Microsoft's FastContext-1.0-4B-RL model, developed by ermiaazarkhalili.

Key Capabilities

  • Function Calling: The model has been specifically fine-tuned on the Salesforce/xlam-function-calling-60k dataset, making it proficient in understanding and generating function calls.
  • Efficient Fine-tuning: Utilizes LoRA (Low-Rank Adaptation) with Unsloth and TRL for efficient supervised fine-tuning, merging adapters directly into the base weights.
  • QLoRA Precision: Trained using 4-bit QLoRA for memory efficiency.

Training Details

The fine-tuning process involved a single epoch with a learning rate of 0.0002 and an effective batch size of 8. The maximum sequence length used during training was 2048 tokens. Training loss observations showed a significant reduction from 1.0659 to 0.1178 over 7,500 steps.

Limitations

  • No Benchmark Evaluation: Currently, no downstream benchmark evaluations have been performed; only training loss is reported.
  • Inherited Biases: The model inherits biases, knowledge cutoff, and potential failure modes from its base model.
  • Specialized Training: Fine-tuned on a single instruction-following dataset, its performance outside this distribution is untested.
  • Merged Adapters: The LoRA adapters are merged into the base weights, meaning the fine-tune cannot be detached.