hongduc05/qwen3-1.7b-vietmedsum-lora-adapter-3

Qwen3-1.7B LoRA for VietMed-Sum

LoRA adapter fine-tuned for Vietnamese medical-conversation summarization.

Base model

Qwen/Qwen3-1.7B

Data preparation

• —Dataset: leduckhai/VietMed-Sum at commit 527a7a496ea30f5d20e5ccb85e59caa7d43b6839
• —Only clean Vietnamese Parquet files were used; English and ASR variants were excluded.
• —Empty records and duplicate normalized transcripts were removed.
• —Summary/transcript word ratio was retained only in [0.05, 0.7].
• —A deterministic random sample of 5,200 clean examples was selected before splitting.
• —Random 8:1:1 split after sampling: {'train': 4160, 'valid': 520, 'test': 520} with seed 42.

LoRA configuration

{ "r": 32, "loraalpha": 64, "targetmodules": [ "qproj", "kproj", "vproj", "oproj" ], "lora_dropout": 0.05, "bias": "none" }

Inference runtime

• —Backend: vLLM offline API
• —Quantization: BitsAndBytes 4-bit (in-flight)

{
  "backend": "vllm",
  "version": "0.10.2",
  "dtype": "float16",
  "quantization": "bitsandbytes",
  "max_model_len": 1536,
  "max_num_seqs": 1,
  "gpu_memory_utilization": 0.85,
  "enforce_eager": true,
  "max_lora_rank": 32,
  "sampling": {
    "temperature": 0.7,
    "top_p": 0.8,
    "top_k": 20,
    "min_p": 0.0,
    "presence_penalty": 1.5,
    "max_tokens": 256,
    "stop_token": "eos",
    "seed": 42
  }
}

Test metrics

• —rouge1_f1: 0.480278
• —rouge2_f1: 0.260241
• —rougeL_f1: 0.412484
• —bleu: 0.242193
• —meteor: 0.405072

Generation latency

• —mean_seconds: 4.419350
• —median_seconds: 3.991283
• —p95_seconds: 7.948480

Metrics are per-example scores averaged over the test set. ROUGE, BLEU and METEOR use Vietnamese word segmentation with Underthesea. Latency measures one vLLM offline LLM.generate request after warm-up on the evaluation GPU; chat-template rendering, input tokenization, metrics and CSV writing are excluded.