shuiyuPyLab/xiaoli-llm-chat
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
52 Commits 1 Branch 4 Tags
6c7c1d38a181d98d432d5c8d87488a8b7d6a1820
Commit Graph
52 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
shuiyu1123 6c7c1d38a1 update data/expanded_distilled.json via distill_expand 2026-07-09 06:44:26 +00:00
XiaLi Bot 3aa8e9d205 XiaLi-Next统一架构: RoPE+GQA+SwiGLU+WeightTying 2026-06-21 04:40:59 +00:00
shuiyu1123 6dbc1405c6 云端训练脚本 + 新词表 + gen_pretrain优化 2026-06-21 02:00:43 +00:00
shuiyu1123 16458aea92 feat: Astron 5并发API + 修复蒸馏P2/P3 + 两阶段训练脚本 + 持续蒸馏守护 
- api_correction.py: 嵌入 Astron key, max_workers=5, 修复 batch_convert/multi_temp 返回空列表bug
- scripts/api_correction.py: 同步更新
- train_max.py: 两阶段训练(预训练+角色对齐), Warmup+Cosine, 分层LR, AMP
- distill_daemon.py: 持续蒸馏守护脚本, 15大类种子话题, 5并发
 2026-06-21 02:00:43 +00:00
shuiyu1123 463cf697fd [自动] 云端一键部署脚本 2026-06-21 01:58:21 +00:00
shuiyu1123 c369804d3c [自动] 云端GPU训练脚本 2026-06-21 01:58:19 +00:00
shuiyu1123 0ccce55ab0 v5: 两阶段生成 - Phase1纯API种子 + Phase2语义扩展 2026-06-20 15:40:47 +00:00
shuiyu1123 c53f4c9049 增加system_prompt和max_tokens参数 - 支持自定义角色设定 2026-06-20 14:47:16 +00:00
shuiyu1123 2bb3b60f18 v4: 双AI互相对话架构 - 2个师生对话对 + 话题生成器 2026-06-20 14:47:06 +00:00
shuiyu1123 70abc8e7ff fix: 同步移除硬编码API key 2026-06-20 13:56:46 +00:00
shuiyu1123 693e7a51e7 fix: 移除硬编码API key, 改为环境变量 2026-06-20 13:56:44 +00:00
shuiyu1123 f8fac5946b feat: 通用知识预训练数据生成器 20大类 2026-06-20 13:52:34 +00:00
shuiyu1123 2ba1ee3416 fix: scripts API同步 2026-06-20 13:24:49 +00:00
shuiyu1123 623a29bafd fix: Astron 5并发 + 修复蒸馏bug 2026-06-20 13:24:47 +00:00
shuiyu1123 1bc19e4c90 feat: 持续蒸馏守护脚本 2026-06-20 13:24:25 +00:00
shuiyu1123 b2c0af564f feat: 两阶段训练脚本 2026-06-20 13:24:24 +00:00
cnb.cEr0mSxFAFA bc6fb0ffe5 remove LFS files for Gitee push (Gitee free tier doesn't support LFS) 2026-06-19 10:50:21 +08:00
cnb.cEr0mSxFAFA 2acde6ae75 chore: enable git LFS for zip and pth files 2026-06-19 10:49:22 +08:00
cnb.cEr0mSxFAFA a88c47e273 feat: v1.3-lite-flash 训练代码 + 新词表(vocab_size=47725) 
- 用 CharTokenizer.build_vocab 从 977K 数据生成词表 (vocab_size=47725)
- train_cnb.py: batch=224 + accum=2 (等效448) + AMP + torch.compile
- build_vocab.py: 词表生成脚本
- vocab_full.json: 含特殊token完整词表
- 模型 41.85M参数, L40 GPU训练
xiaoli-llm-v1.3-lite-flash 		2026-06-19 10:34:37 +08:00
shuiyu1123 a0047a012d revert: restore original Dockerfile that was working 2026-06-18 20:36:04 +00:00
shuiyu1123 7b47faac50 fix: add vscode-server dir to Dockerfile, use vscode event for SSH 2026-06-18 20:18:20 +00:00
shuiyu1123 c6ff7be55a chore: revert to version 3 tasks format for SSH support 2026-06-18 20:05:41 +00:00
shuiyu1123 f5dd3e5874 chore: add vscode service to enable SSH in workspace 2026-06-18 19:53:09 +00:00
shuiyu1123 ac07c7f06a chore: revert to workspace format (version 3) 2026-06-18 19:22:03 +00:00
shuiyu1123 230d47cade fix: CI pipeline clones repo to /workspace directly, not subdir 2026-06-18 19:09:10 +00:00
shuiyu1123 5732229ed5 chore: add debug echo to pipeline script 2026-06-18 19:06:17 +00:00
shuiyu1123 ffb5825407 fix: correct vocab_size to max_id+1 (was len(vocab)) 2026-06-18 18:43:47 +00:00
shuiyu1123 a02401aa75 fix: LLMTrainer supports lr param 2026-06-18 18:40:25 +00:00
shuiyu1123 edf3a83e3c chore: rewrite .cnb.yml to CI pipeline format with api_trigger event 2026-06-18 18:27:38 +00:00
shuiyu1123 9799aa00c5 chore: set training epochs to 100 2026-06-18 18:25:44 +00:00
shuiyu1123 456c443e9c feat: add CNB cloud training config (GPU Dockerfile + .cnb.yml + train_cnb.py) 2026-06-18 18:09:59 +00:00
shuiyu1123 eb7717e9ce feat: add new general distill script, merged dataset v3 (977K pairs), legacy vocab_v2 2026-06-18 17:56:40 +00:00
shuiyu1123 231af72de9 大整理: 仓库重组 + 新词表 + 新README 
- 清理仓库: 删除所有废弃的generate_*脚本和旧版数据文件
- 目录重组:
  - data/       → 训练数据
  - scripts/    → 数据生成脚本
  - templates/  → Web UI
- 重命名:
  - train_v2.py  → train.py
  - vocab_v3.json → vocab.json (BPE, 6.8k tokens)
- 更新 train.py 指向新词表路径
- 重写 README.md 完整文档
- 更新 .gitignore
v0.1-epoch7 		2026-06-18 11:17:30 +00:00
shuiyu1123 6e0ced4772 添加v3词表 (BPE, 6.8k tokens, 666k数据训练) 
- 从666k条训练数据自动学习
- 5812单字 + 1000高频词组
- 编码效率提升33%
- 含小狸特色词组: 捉迷藏/冰淇淋/为什么等
 2026-06-17 14:31:25 +00:00
shuiyu1123 aead65192e 整理仓库: 清理旧分片/测试文件, 更新.gitignore, 添加训练数据 2026-06-16 14:12:33 +00:00
shuiyu1123 075492e391 添加训练数据 (666k条, 2分片) 2026-06-16 14:01:47 +00:00
shuiyu1123 8efb4faf73 添加合并脚本: bash merge_model.sh 2026-06-16 12:48:39 +00:00
shuiyu1123 2976bf2030 添加模型文件 (Epoch 7, 分2片, bash merge_model.sh合并) 2026-06-16 12:47:34 +00:00
shuiyu1123 9cef1d9ba1 添加模型文件分片 (Epoch 7, 3 parts + merge script) 2026-06-16 12:10:24 +00:00
shuiyu1123 91db488ad4 同步代码更新 (2026-06-16) 
- 移除 generate_diverse_answers.py（多样化回答已废止）
- 更新 generate_fresh_qa.py（目标50万条新问答）
- 更新 generate_multiturn.py（6~12轮长对话，目标50万）
- 更新 regenerate_all_high_temp.py（独立保存，不追加训练集）
- 更新 .gitignore（忽略新输出目录）
- 所有API Key改为环境变量读取
 2026-06-16 11:48:30 +00:00
shuiyu1123 eb1d62f85c 紧急修复：移除所有硬编码API Key，改为环境变量 ASTRON_KEY 2026-06-13 04:04:02 +00:00
shuiyu1123 7f27812efd 更新README：旨在最小设备上运行的闲聊心理治疗模型 2026-06-12 15:16:56 +00:00
shuiyu1123 d962c768a5 同步：更新daily_data_gen.py，200+话题+3000条/天 (2026-06-12) 2026-06-12 15:12:40 +00:00
shuiyu1123 7f8f607f26 同步当前架构和数据 (2026-06-11) 
- 训练脚本: train_v2.py + llm_chat_improved.py
- API生成: api_correction.py
- 高温多样化: regenerate_all_high_temp.py
- 每日数据生成: daily_data_gen.py
- 数据: xiaoli_train_v2.json(40776条) + vocab_v2.json
- 蒸馏问题: exam_distill_questions等
- Web服务: web_server.py + chat.html
- 新增 .gitignore 排除旧版文件
 2026-06-11 14:07:52 +00:00
shuiyu1123 4df3254569 add LICENSE. 
Signed-off-by: shuiyu1123 <f17376088816@163.com>
 2026-06-06 00:26:36 +00:00
shuiyu1123 d2aa6013d7 上传 templates/chat.html v1.0.0 v1.1 2026-06-05 15:09:00 +00:00
shuiyu1123 b55d686d3a 上传 xiaoli_data.json 2026-06-05 15:08:47 +00:00
shuiyu1123 f9594de862 上传 multiturn_data.json 2026-06-05 15:08:46 +00:00
shuiyu1123 1fb1a0393a 上传 all_data_cleaned.json 2026-06-05 15:08:44 +00:00
shuiyu1123 f7e61f26d1 上传 llm_chat_improved.py 2026-06-05 15:08:42 +00:00