ermiaazarkhalili/Qwen3.5-2B-SFT-Fable5-Glint

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ermiaazarkhalili/Qwen3.5-2B-SFT-Fable5-Glint is a 2.3 billion parameter Qwen3.5-based language model, fine-tuned using LoRA on a private instruction-following dataset. Developed by ermiaazarkhalili, this model demonstrates improved next-token accuracy over its base model on held-out data. It is designed for instruction-following tasks, leveraging a 4096 token context length and 4-bit QLoRA precision.

Loading preview...

Model Overview

This model, ermiaazarkhalili/Qwen3.5-2B-SFT-Fable5-Glint, is a 2.3 billion parameter language model based on the unsloth/Qwen3.5-2B architecture. It has been supervised fine-tuned using LoRA (Low-Rank Adaptation) on a private dataset, ermiaazarkhalili/Fable-5-Glint-Clean, to enhance its instruction-following capabilities.

Key Characteristics

  • Base Model: unsloth/Qwen3.5-2B
  • Parameters: 2.3 billion
  • Fine-tuning Method: LoRA supervised fine-tuning via Unsloth and TRL, with a LoRA rank of 16 and alpha of 16.
  • Precision: Trained using 4-bit QLoRA.
  • Context Length: Supports a maximum sequence length of 4096 tokens.

Performance and Evaluation

On a held-out split of the Fable-5-Glint-Clean dataset, this fine-tuned model shows a significant improvement in next-token accuracy compared to its base model:

  • Top-1 Accuracy: Increased by +0.0900 (from 0.5970 to 0.6869).
  • Top-5 Accuracy: Increased by +0.0692 (from 0.8383 to 0.9075).

Limitations

  • No external benchmark evaluations have been conducted; reported metrics are based on training loss observations and held-out accuracy.
  • Inherits biases and limitations from the base model.
  • Fine-tuned on a single instruction-following dataset, meaning performance outside this distribution is untested.
  • LoRA adapters are merged, preventing detachment from this fine-tune.