qwen3 seq_cls

[vibe-viscot]

Describe the bug
What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图)
qwen3 seq_cls分类有问题，训练出来的eval_acc很低，qwen3-base模型是正常的，可能是chat_template的原因？

训练结果如下：
qwen3-0.6b
{"eval_loss": 1.90731037, "eval_acc": 0.6, "eval_runtime": 3.5113, "eval_samples_per_second": 37.023, "eval_steps_per_second": 37.023, "epoch": 0.9995338, "global_step/max_steps": "804/804", "percentage": "100.00%", "elapsed_time": "26m 25s", "remaining_time": "0s"}
qwen3-0.6b-base
{"eval_loss": 1.30270481, "eval_acc": 0.74615385, "eval_runtime": 3.6257, "eval_samples_per_second": 35.855, "eval_steps_per_second": 35.855, "epoch": 0.9995338, "global_step/max_steps": "804/804", "percentage": "100.00%", "elapsed_time": "25m 59s", "remaining_time": "0s"}

qwen2.5-0.5b-instruct
{"eval_loss": 1.38452613, "eval_acc": 0.73076923, "eval_runtime": 2.8806, "eval_samples_per_second": 45.13, "eval_steps_per_second": 45.13, "epoch": 0.9995338, "global_step/max_steps": "804/804", "percentage": "100.00%", "elapsed_time": "20m 56s", "remaining_time": "0s"}
qwen2.5-0.5b-base
{"eval_loss": 1.33685374, "eval_acc": 0.7, "eval_runtime": 2.7394, "eval_samples_per_second": 47.455, "eval_steps_per_second": 47.455, "epoch": 0.9995338, "global_step/max_steps": "804/804", "percentage": "100.00%", "elapsed_time": "20m 53s", "remaining_time": "0s"}

Your hardware and system info
Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)

Additional context
Add any other context about the problem here(在这里补充其他信息)