gary109/qwen3-vl-4b-instruct-sft-seed3407-20251231-0201-merged_16bit

qwen3-vl-4b-instruct-sft-seed3407-20251231-0201

簡介

• 任務: qwen3-vl-sft
• Base model: unsloth/Qwen3-VL-4B-Instruct
• 作者: 蓋瑞王
• 團隊: UNIFORM
• 訓練設定來源: /workspace/project/unsloth/configs/qwen3-vl-4b-vision-sft.yaml
• 輸出目錄: outputs/qwen3-vl-4b-instruct-sft-seed3407-20251231-0201

訓練設定（節錄）

seed: 3407
model:
  name: unsloth/Qwen3-VL-4B-Instruct
  max_seq_length: 2048
  load_in_4bit: true
  fast_inference: false
  use_gradient_checkpointing: unsloth
peft:
  lora_rank: 16
  lora_alpha: 16
  lora_dropout: 0
  finetune_vision_layers: true
  finetune_language_layers: true
  finetune_attention_modules: true
  finetune_mlp_modules: true
  use_rslora: false
data:
  dataset_name: unsloth/LaTeX_OCR
  split: train
  instruction: Write the LaTeX representation for this image.
  image_field: image
  text_field: text
sft:
  per_device_train_batch_size: 1
  gradient_accumulation_steps: 4
  max_steps: 1000
  learning_rate: 0.0002
  warmup_steps: 5
  optim: adamw_8bit
  weight_decay: 0.001
  lr_scheduler_type: linear
eval:
  enabled: true
  dataset_name: unsloth/LaTeX_OCR
  split: test
  max_samples: 200
  max_new_tokens: 128
  temperature: 0.0
  top_p: 1.0
  do_sample: false

評測摘要

• 評測摘要來源: outputs/qwen3-vl-4b-instruct-sft-seed3407-20251231-0201/best_checkpoints.json
• 最佳 checkpoint: checkpoint-200
• num_samples: 200
• exact_match: 0.35
• edit_distance: 15.835
• eval.split: test
• eval.max_samples: 200

評測摘要（表格）

eval_results.json

source	num_samples	exact_match	edit_distance
eval_results.json	200	0.3300	15.8550

best_checkpoints.json

rank	checkpoint	exact_match	edit_distance	note
1	checkpoint-200	0.3500	15.8350	-
2	checkpoint-100	0.3450	16.7500	-
3	checkpoint-400	0.3300	15.8550	-

備註

• 請補上授權、用途說明與限制等資訊。