 ViperEkura
|
598e1ce4ae
|
refactor: 重构打包模块，新增 BFD/FFD/Greedy 三种 bin-packing 算法，默认 BFD
- 将 pipeline/packing.py 拆分为 packing/ 子包 (base/stream/binpack)
- 新增 BfdPacker(默认)/FfDPacker/GreedyPacker，移除 StreamingPacker
- 超长序列直接截断至 pack_size
- group_size 语义改为"每 N 个 chunk 合并为一块"，默认 1000
- 新增 AutoTokenizer.token_to_id()，修复 ChatML 中 hacky 的 nl_id 获取
- pad_value 默认改为 2(pad_token_id)，position_ids pad=0, loss_mask pad=False
- 新增 position_ids 打包后归零一致性测试
- scripts/cache_h5.py 新增 --pack-algo 参数
|
2026-07-03 16:17:27 +08:00 |
|
|
ViperEkura
|
2f919e9243
|
feat: add MetaMathQA SFT export script and fix ChatML newline encoding
|
2026-07-02 17:40:03 +08:00 |
|
|
ViperEkura
|
fd99b9d852
|
feat: add OpenHermes-2.5 SFT export script
|
2026-06-04 14:27:04 +08:00 |
|
 jiachun
|
efba7a009f
|
feat: add sft scripts for Magicoder-Evol-Instruct-110K and alpaca-gpt4-data
- sft_magicoder.py: export ise-uiuc/Magicoder-Evol-Instruct-110K
- sft_alpaca_gpt4.py: export llm-wizard/alpaca-gpt4-data with instruction+input
- remove deprecated SFT scripts
|
2026-06-01 15:50:13 +08:00 |
|
|
ViperEkura
|
625695fd71
|
fix: padding 默认值改为 0，避免 loss_mask 误计 padding 位置
|
2026-05-15 16:59:42 +08:00 |
|
|
ViperEkura
|
cb6bfcb976
|
refactor: 重构流水线架构，添加Pipeline抽象并拆分IOHandler
|
2026-04-23 19:45:57 +08:00 |
|
|
ViperEkura
|
a38334f4ce
|
refactor: 修改模型架构
|
2026-04-16 18:57:21 +08:00 |
|
|
ViperEkura
|
f44ad6912e
|
fix: 修复特殊token 问题
|
2026-04-02 16:16:02 +08:00 |
|
|
ViperEkura
|
35963bcb08
|
reafactor: 重构项目
|
2026-03-30 20:58:51 +08:00 |
|
|
ViperEkura
|
07ef471aa6
|
chore: 增加 h5_cache 脚本
|
2026-03-21 21:33:38 +08:00 |
|
|
ViperEkura
|
909f1b400c
|
refactor: 修改项目结构
|
2026-03-21 21:25:59 +08:00 |
|