Qwen3 Chat Summarization

Training

• Requested and saved base model: Qwen/Qwen3-1.7B
• Unsloth-resolved ID before portable normalization: Qwen/Qwen3-1.7B
• Unsloth FP16/BF16 LoRA training, not QLoRA.
• No quantization was used during training.
• Winner run: cf4.
• LoRA rank: 32; alpha: 128; dropout: 0.03.
• Target modules: q_proj, k_proj, v_proj, o_proj.
• Learning rate: 0.0004; minimum LR ratio: 0.05; optimizer: adamw_torch_fused.
• Per-GPU batch size: 8; gradient accumulation: 16; effective batch size: 128.
• Maximum epochs: 5; stopped epoch: 4.000; weight decay: 0.01.
• Best validation loss: 0.969759; perplexity: 2.637308.
• Baseline cf4 validation loss: 0.969759; absolute improvement: 0.000000; improvement >= 0.01: False.
• Context length: 1024; max new tokens: 70.
• Prompt uses a Vietnamese system instruction and a manual An/Bình/Chi/Dũng one-shot.
• Thinking is disabled with enable_thinking=False.
• Data split: grouped by normalized Chat with seed 42; exactly 300 validation rows and 300 test rows, with all remaining rows used for training; workbook rows after cleaning: 9282.

Prompt

Inference

• vLLM loads Qwen/Qwen3-1.7B in float16; BitsAndBytes is intentionally disabled to avoid duplicate custom-op registration in the pinned vLLM stack.
• The best PEFT LoRA adapter is attached with LoRARequest; the base model is not merged or uploaded.
• Sampling: temperature=0.7, top_p=0.8, top_k=20, min_p=0, max_tokens=70, seed=42.
• Metrics tokenize Vietnamese text with underthesea.
• METEOR uses exact Vietnamese token matches without English stemming or WordNet.

Evaluation on held-out test split