guanghulab/brain/training-status.json

75 lines
3.3 KiB
JSON
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

{
"version": "1.2",
"last_updated": "2026-05-18T22:45+08:00",
"updated_by": "铸渊 (ICE-GL-ZY001)",
"mother_model": {
"status": "completed",
"status_label": "✅ 已完成COS文件已修复 generation_config.json",
"model": "Qwen/Qwen2.5-7B",
"final_loss": 0.2638,
"final_epoch": 3,
"total_epochs": 3,
"cos_path": "cos://sy-finetune-corpus-1317346199/models/qwen25-7b-sft/final/",
"completed_at": "2026-05-18T~06:00+08:00",
"fixes": {
"generation_config_eos": "151643→151645 ✅ 2026-05-18 修复根因model.generate()读取generation_config而非config",
"default_system": "已清除 ✅"
}
},
"code_model": {
"status": "completed",
"status_label": "✅ 已完成train_coder.log确认DONE!",
"model": "Qwen/Qwen2.5-Coder-7B",
"cos_path": "cos://sy-finetune-corpus-1317346199/models/qwen25-coder-7b-sft/final/",
"fixes": {
"generation_config_eos": "151643→151645 ✅ 2026-05-18 修复"
}
},
"distill_shuangyan_15b": {
"status": "running",
"status_label": "🔄 蒸馏训练中Screen 87779, step=50 loss=6.68, Epoch 1/3",
"version": "v7",
"script": "scripts/distill_mother.py",
"teacher": "母模型7B SFT",
"student": "Qwen2.5-1.5B-Instruct",
"method": "KL散度(TEMP=2.0, ALPHA=0.7) + SFT混合",
"host": "connect.westd.seetacloud.com:23647",
"ssh_password": "HkM43lFVUIsc",
"started_at": "2026-05-18T22:41+08:00",
"estimated_completion": "~01:00+08:00",
"cos_output": "models/shuangyan-15b-distill/final/",
"fixes": [
"v7: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True + empty_cache在epoch转换/模型加载/每500步",
"v6: vocab_size截断 fix — Teacher(7B)=152064 → Student(1.5B)=151936",
"v5: HuggingFace forward参数名修正 (ii→input_ids, am→attention_mask)",
"v4: num_workers=0 (避免CUDA fork)",
"v5: 移除teacher_logits预存储改为即时forward"
]
},
"pipeline": {
"steps": [
{"name": "母模型SFT训练", "status": "completed", "label": "✅ 完成 (loss=0.2638)"},
{"name": "上传COS (母模型)", "status": "completed", "label": "✅ 已完成"},
{"name": "代码模型SFT训练", "status": "completed", "label": "✅ 已完成"},
{"name": "上传COS (代码模型)", "status": "completed", "label": "✅ 已完成"},
{"name": "generation_config修复", "status": "completed", "label": "✅ 两个模型已修复"},
{"name": "霜砚1.5B蒸馏", "status": "running", "label": "🔄 训练中Screen 87779, Epoch 1/3"},
{"name": "上传COS (蒸馏输出)", "status": "pending", "label": "⚪ 等待蒸馏完成"}
]
},
"next_phases": {
"distill_shuangyan_15b": "🔄 训练中,预计~01:00完成",
"shuangyan_deep_sft": "⏳ 等待蒸馏完成后用霜砚语料深度微调",
"distill_coder_to_zhuyuan_15b": "⏳ 等待Phase 1+2完成",
"zhuyuan_deep_sft": "⏳ 等待Phase 3完成"
},
"alerts": [
"⚠️ 之前误判代码模型已完成COS上有历史checkpoint-540",
"⚠️ 蒸馏经过7次迭代才稳定路径错误→存储爆炸→CUDA fork→参数名错误→vocab不匹配→OOM(×2)",
"⚠️ v7 所有已知崩溃已修复,预计~01:00完成"
],
"links": {
"notion": "https://www.notion.so/363fb92f383181029fafef8438d6c300"
}
}