diff --git a/brain/training-status.json b/brain/training-status.json index 1016b27..dc15ff7 100644 --- a/brain/training-status.json +++ b/brain/training-status.json @@ -1,74 +1,76 @@ { - "version": "1.2", - "last_updated": "2026-05-18T22:45+08:00", + "version": "1.3", + "last_updated": "2026-05-20T09:48+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)", + "status_label": "✅ 已完成 · 已上传COS bingshuo-1317346199", "model": "Qwen/Qwen2.5-7B", - "final_loss": 0.2638, + "final_loss": 1.5229, "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", + "cos_path": "cos://bingshuo-1317346199/models/qwen25-7b-sft/final/", + "completed_at": "2026-05-19T~22:30+08:00", "fixes": { - "generation_config_eos": "151643→151645 ✅ 2026-05-18 修复(根因:model.generate()读取generation_config而非config)", + "generation_config_eos": "151645 ✅", "default_system": "已清除 ✅" } }, "code_model": { "status": "completed", - "status_label": "✅ 已完成(train_coder.log确认DONE!)", + "status_label": "✅ 已完成 · 已上传COS bingshuo-1317346199", "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" - } + "final_loss": 1.822, + "cos_path": "cos://bingshuo-1317346199/models/qwen25-coder-7b-sft/final/", + "fixes": {} }, "distill_shuangyan_15b": { "status": "running", - "status_label": "🔄 蒸馏训练中(Screen 87779, step=50 loss=6.68, Epoch 1/3)", + "status_label": "🔴 蒸馏训练中 · Ep1/3 step~4850/15781 loss≈1.96", "version": "v7", - "script": "scripts/distill_mother.py", - "teacher": "母模型7B SFT", + "script": "scripts/distill_mother_v7.py", + "teacher": "母模型7B SFT (COS bingshuo)", "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" - ] + "params": "B=2, GA=16, LR=1e-5, TEMP=2.0, ALPHA=0.7", + "host": "connect.westd.seetacloud.com:23647 (RTX PRO 6000 96GB)", + "started_at": "2026-05-20T09:23+08:00", + "estimated_completion": "~12:00~13:00+08:00 (3 epoch, ~45min/ep)", + "cos_output": "models/shuangyan-15b-distill/", + "pipeline_watchdog": "/tmp/distill_pipeline.sh (PID 581561, 自动触发Track2)" + }, + "distill_coder_15b": { + "status": "pending", + "status_label": "⏳ 等待Track1完成 + 管道自动触发", + "script": "scripts/distill_coder_v2.py", + "teacher": "代码模型7B SFT (COS bingshuo)", + "student": "Qwen2.5-Coder-1.5B-Instruct", + "params": "B=2, GA=16, LR=1e-5 (预修OOM)", + "cos_output": "models/qwen25-15b-coder-distill/" }, "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"} + {"name": "母模型SFT训练+上传COS", "status": "completed", "label": "✅"}, + {"name": "代码模型SFT训练+上传COS", "status": "completed", "label": "✅"}, + {"name": "Track1 母模型→霜砚1.5B蒸馏", "status": "running", "label": "🔄 Ep1/3 step~4850/15781"}, + {"name": "Track1 上传COS", "status": "pending", "label": "⏳ 蒸馏自带上传"}, + {"name": "Track2 代码模型→铸渊1.5B蒸馏", "status": "pending", "label": "⏳ 管道自动触发"}, + {"name": "Track2 上传COS", "status": "pending", "label": "⏳"} ] }, - "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" + "website_status": { + "login_api": "✅ 正常", + "training_panel": "✅ 实时显示 (fetch_train.py v4)", + "js_all_functions": "✅ 已修复 (setInterval恢复)", + "gpu_monitoring": "✅ training-status.json每分钟更新" }, "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" + "⚠️ 本次修复了fetch_train.py v3的3个bug(phase检测/epoch提取/tail长度)", + "⚠️ 前一次index.html修复(3583413)引入语法错误→全站JS崩溃→已修复", + "⚠️ Track2的OOM已在Track1完成前预修", + "⚠️ 冰朔强调的工作流程:先读仓库记忆→理解全局→再动手" ], "links": { - "notion": "https://www.notion.so/363fb92f383181029fafef8438d6c300" + "notion": "https://www.notion.so/363fb92f383181029fafef8438d6c300", + "repo": "https://guanghulab.com/code/bingshuo/guanghulab" } }