ermiaazarkhalili/Qwen3.5-4B-SFT-Fable5-Glint

Model Overview

ermiaazarkhalili/Qwen3.5-4B-SFT-Fable5-Glint is a 4.7 billion parameter language model, developed by ermiaazarkhalili. It is a LoRA (Low-Rank Adaptation) fine-tune of the unsloth/Qwen3.5-4B base model, specifically supervised fine-tuned on the private ermiaazarkhalili/Fable-5-Glint-Clean dataset. The fine-tuning process utilized Unsloth and TRL libraries, employing QLoRA with 4-bit precision and a maximum sequence length of 4096 tokens.

Key Capabilities and Performance

This model is primarily designed for instruction-following tasks, benefiting from its supervised fine-tuning. Evaluation on a held-out split of the training data shows a notable improvement in next-token accuracy compared to the base model:

• Top-1 accuracy: Increased from 0.6312 (base) to 0.7226 (this model), a delta of +0.0914.
• Top-5 accuracy: Increased from 0.8715 (base) to 0.9318 (this model), a delta of +0.0603.

These metrics indicate enhanced performance on tasks similar to its training distribution. The model inherits the apache-2.0 license from its base.

Limitations and Considerations

It is important to note that no external benchmark evaluations have been conducted on this checkpoint; reported numbers are based on training loss observations and held-out accuracy on the specific training dataset. The model inherits the biases, knowledge cutoff, and potential failure modes of its unsloth/Qwen3.5-4B base. Its behavior outside the distribution of the Fable-5-Glint-Clean dataset is untested. Due to the merging of LoRA adapters, the fine-tuned model cannot be detached from its base weights.