ChenxiZeng/interrupt-detection-cot-lora

interrupt_detection_cot_lora

This model is a fine-tuned version of Qwen/Qwen3-0.6B on the interrupt_detection_cot dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

• Training set: interrupt_detection_cot_train.jsonl (4000 samples)
• Validation set: interrupt_detection_cot_val.jsonl (500 samples)
• Test set: interrupt_detection_cot_test.jsonl (500 samples)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 3.0

Training results

Metric	Value
Train Loss (final, epoch avg.)	0.5221
Eval Loss (final @ step 939)	0.638
Best Eval Loss (@ step 700)	0.4512
Eval frequency	every 100 steps (eval_steps=100)
Total Training Steps	939
Training Runtime	2899.46s (~48.3 min)

• Best model checkpoint: checkpoint-700 (lowest validation loss; use this for deployment when early stopping is enabled).
• Early stopping: after the best eval step, validation loss rises while training loss keeps bouncing down — a sign of overfitting; stop once eval fails to improve for patience evaluations.

training_loss.png is generated with raw logged train loss (noisy batches) and EMA-smoothed train loss on the same axes; see regenerate_loss_figure.py to reproduce.

Framework versions

• PEFT 0.15.2
• Transformers 4.52.4
• Pytorch 2.10.0+cu128
• Datasets 3.6.0
• Tokenizers 0.21.1