CounselTrajLM-Qwen3-8B / training_summary.json
ZhiyangQi97's picture
Release merged BF16 CounselTrajLM based on Qwen3-8B
4a13f72 verified
Raw History Blame Contribute Delete
876 Bytes
{
"model": "CounselTrajLM-Qwen3-8B",
"base_model": "Qwen/Qwen3-8B",
"base_revision": "b968826d9c46dd6066d109eabc6255188de91218",
"training_dataset": "CounselTrajD",
"split": "dialogue-disjoint 90/10 train/validation",
"train_examples": 425679,
"validation_examples": 47528,
"fine_tuning_method": "QLoRA",
"release_format": "adapter merged into unquantized BF16 base weights",
"learning_rate": 0.0002,
"epochs": 3,
"checkpoint_policy": "final_epoch",
"validation_role": "monitoring",
"lr_scheduler": "cosine",
"warmup_ratio": 0.06,
"effective_batch_size": 16,
"max_sequence_length": 2048,
"max_history_characters": 1800,
"seed": 20260704,
"global_step": 79815,
"thinking_enabled": false,
"lora": {
"rank": 32,
"alpha": 64,
"dropout": 0.05
},
"train_loss": 1.3573430852992174,
"validation_loss": 1.513865351676941
}