qwen3-0.6b-sft

Qwen3-0.6B SFT Adapter

This is a Supervised Fine-Tuned (SFT) LoRA adapter for the Qwen/Qwen3-0.6B base model. It was trained to improve instruction-following capabilities and stylistic formatting.

Model Details

Model Description

This adapter applies Low-Rank Adaptation (LoRA) to the Qwen3-0.6B model to align it with human instructional prompts.

• Developed by: Agha Salik Ali and Uzair Nadeem
• Model type: LoRA Adapter for Causal Language Model
• Language(s) (NLP): English
• Finetuned from model: Qwen/Qwen3-0.6B

Uses

Direct Use

This adapter is intended to be loaded on top of the base Qwen3-0.6B model for conversational interactions, natural language generation, and instruction following.

Out-of-Scope Use

Due to the constraints of the 0.6B parameter base model, this fine-tuned version is susceptible to persistent hallucinations and format collapse (e.g., infinite repetitive loops) on complex creative tasks. It should not be relied upon for factual accuracy without an external retrieval mechanism.

Training Details

Training Data

The model was fine-tuned on the vicgalle/alpaca-gpt4 dataset, utilizing the standard Alpaca chat template. A 90/10 train-evaluation split was used.

Training Procedure

The model was trained using the SFTTrainer from the trl library with the base model's native thinking tokens disabled (enable_thinking=False).

Training Hyperparameters

• Training regime: fp16 mixed precision
• LoRA Rank (r): 64
• LoRA Alpha: 128
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• LoRA Dropout: 0.1
• Learning Rate: 5e-05
• Train Batch Size: 2
• Gradient Accumulation Steps: 8
• Number of Epochs: 2

Compute Infrastructure

• Hardware: Kaggle Dual NVIDIA T4 GPUs

Evaluation

Testing Data & Metrics

The model was evaluated using a customized set of 10 prompts to measure semantic alignment and surface-level string matching.

• Metrics: BLEU, BERTScore F1

Results

• Validation Loss: 1.2782
• BLEU: 5.3696
• BERTScore F1: 0.8529

Framework versions

• PEFT 0.18.1
• Transformers
• TRL