github
/
minimind
mirror of https://github.com/jingyaogong/minimind.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Commit Graph

  • 465f87c8c5
         Merge 695aa1fdb6 into f659b55761 liyang 2026-10-25 23:59:50 +0800
  • 695aa1fdb6 feat(lora): make LoRA rank configurable in train_lora.py and eval_llm.py BBqBuleberry 2026-10-25 23:22:52 +0800
  • 055e1c08e4 docs: add project overview and design notes Hwe1916 2026-10-24 00:45:28 +0800
  • 57ffce270f feat(data): add small benchmark samples Hwe1916 2026-10-24 00:45:28 +0800
  • cd6efc9206 test: add 3GPP agent unit tests Hwe1916 2026-10-24 00:45:28 +0800
  • 4d960838e4 feat(eval): add retrieval and benchmark evaluation Hwe1916 2026-10-24 00:45:27 +0800
  • ee3fc66803 feat(train): add QLoRA training scripts Hwe1916 2026-10-24 00:45:27 +0800
  • cb26da0695 feat(data): add 3GPP corpus preparation scripts Hwe1916 2026-10-24 00:45:27 +0800
  • eebc926c51 feat(web): add demo frontends for the RAG agent Hwe1916 2026-10-24 00:45:27 +0800
  • e83b906ed0 feat(agent): add 3GPP RAN2 proposal RAG agent Hwe1916 2026-10-24 00:45:27 +0800
  • bc56be6495 chore(repo): ignore artifacts and generated datasets Hwe1916 2026-10-24 00:45:27 +0800
  • eefa663dd4 style(model): silence missing-import type warning Hwe1916 2026-10-24 00:45:27 +0800
  • 9a25cceaeb fix(deps): bump ujson to 5.10.0 for Python 3.12 Hwe1916 2026-10-24 00:45:27 +0800
  • f9e61ee0e8
         Merge e4872e6898 into f659b55761 Scalwayssilence 2026-10-19 17:28:54 +0800
  • e4872e6898
         test: cover padded SDPA outputs gradients and fallback paths Scalwayssilence 2026-10-19 17:27:01 +0800
  • c401918813
         fix: use SDPA for padded attention without a KV cache Scalwayssilence 2026-10-19 17:24:42 +0800
  • 91d2ef159b
         Merge cd36b4a013 into f659b55761 Youngju KIM 2026-10-17 03:58:09 +0900
  • cd36b4a013
         [fix] decide native-vs-transformers loading by directory name, not substring YoungjuKim 2026-10-17 03:53:56 +0900
  • 9beae2da75
         Merge a067e2512c into f659b55761 Dkkt 2026-10-15 23:55:17 +0800
  • a067e2512c [fix] SFTDataset: keep assistant span when long prompt is truncated uphomer 2026-10-15 22:26:21 +0800
  • 9f4440701b
         Merge 7bde9752c1 into f659b55761 yi111 2026-10-09 23:12:39 +0800
  • 7bde9752c1 docs: clarify how to read the GRPO policy loss curve Yi-111-a 2026-10-09 23:12:13 +0800
  • c0554088a3
         Merge 86aad50621 into f659b55761 cyanseek 2026-10-02 19:54:58 +0800
  • 86aad50621 [fix] Keep streamed tool calls out of response content cyanseek 2026-10-02 19:54:21 +0800
  • 2884eeb05f
         Merge dcbd31bf3d into f659b55761 yi111 2026-10-01 11:31:00 +0800
  • dcbd31bf3d [fix] keep malformed tool calls from crashing agent RL training Yi-111-a 2026-10-01 11:30:27 +0800
  • 04203cc1ba
         Merge 88031f1ac1 into f659b55761 LiberServer 2026-09-29 17:05:14 +0800
  • 88031f1ac1 [fix] install plugin with uv in vllm image Draic 2026-09-29 16:30:37 +0800
  • 92ff635378 [ci] build image from clean branch Draic 2026-09-29 16:23:42 +0800
  • 1a636c7df8 [fix] ignore large model and dataset artifacts Draic 2026-09-29 16:20:56 +0800
  • 142eb5905d [update] add GHCR build workflow Draic 2026-09-29 16:04:33 +0800
  • b3bd25fb19
         Merge 9cb07dd1ec into f659b55761 Neil Ge 2026-09-29 16:15:42 +0800
  • e1c4c2030b Add server RLVR experiment results qi53 2026-09-28 11:40:50 +0800
  • 6163c89581 Add RLVR training dynamics audit Your Name 2026-09-28 11:15:22 +0800
  • 63d8da298c
         Merge 38028f4811 into f659b55761 Momoyeyu 2026-09-28 00:33:50 +0800
  • 38028f4811
         [update] lock uv dependencies for reproducible installs momoyeyu 2026-09-28 00:32:54 +0800
  • f4d34d38e7
         Merge 743420daac into f659b55761 1Dreamer666 2026-09-26 18:04:15 +0800
  • 771192f0a2
         Merge 138c367ddb into f659b55761 Yixian Chen 2026-09-26 03:01:12 -0700
  • 62cb5dbc5c [fix] ujson 在新版 Python 下编译失败阻塞依赖安装，移除该未使用依赖并补充 pyproject.toml momoyeyu 2026-09-26 09:48:00 +0800
  • 641d956015 [update] 新增一键训练 pipeline 脚本 run_pipeline.sh momoyeyu 2026-09-26 11:34:18 +0800
  • 00da95959f
         Merge 655d4705eb into f659b55761 Corey52HZ 2026-09-24 23:33:15 +0800
  • 76de934420
         Merge f81a948b52 into f659b55761 林SO 2026-09-24 16:53:32 +0800
  • b8593bae79
         Merge 56a21ed3c2 into f659b55761 AmirX0 2026-09-23 23:54:52 +0800
  • fce2e7f1b7
         Merge 0761ffa7cd into f659b55761 王翛玄 2026-09-23 17:04:35 +0800
  • 62235b2c43
         Merge 66aef88185 into f659b55761 AmirX0 2026-09-23 17:01:01 +0800
  • 6c65a79e9a
         Merge ccb352f901 into f659b55761 Dujus 2026-09-23 04:42:04 -0400
  • 1855c78bfa
         Merge f82990b0ed into f659b55761 HITSZJiahaoLiu 2026-09-23 09:35:51 +0800
  • 9ab410c71e
         Merge 15b2dd4df4 into f659b55761 HITSZJiahaoLiu 2026-09-23 09:34:44 +0800
  • 87e3aa1d5a
         Merge 1fe7e5a935 into f659b55761 Daoyuan Li 2026-09-22 17:16:28 -0700
  • 1fe7e5a935 fix: unpack MoE experts when converting back to PyTorch Daoyuan Li 2026-09-22 17:16:13 -0700
  • 98c62c7996
         Merge 437eae89d6 into f659b55761 Jiyang 2026-09-23 07:32:02 +0800
  • f659b55761 [fix] residual step order master jingyaogong 2026-09-23 00:07:30 +0800
  • 6fff76fe04
         Merge pull request #868 from DaoyuanLi2816/fix/final-accumulation-checkpoint jingyaogong 2026-09-22 23:39:01 +0800
  • 7da0f92ff6
         Merge pull request #867 from DaoyuanLi2816/fix/lora-inference-paths jingyaogong 2026-09-22 23:20:45 +0800
  • f82990b0ed feat: add multi-turn tool-use on-policy distillation HITSZJiahaoLiu 2026-09-22 23:17:32 +0800
  • 15b2dd4df4 fix(server): complete streaming chunks and stop after generation errors HITSZJiahaoLiu 2026-09-22 20:51:41 +0800
  • a6ffddc0ea fix: update residual gradients before saving final checkpoints Daoyuan Li 2026-09-21 23:24:55 -0700
  • 4b86522f8a fix: align LoRA inference paths with training output Daoyuan Li 2026-09-21 23:24:55 -0700
  • f81a948b52 [feat] get_lr 增加可选 warmup_frac 参数（默认关闭，行为逐值不变） Linxiushen 2026-09-20 21:41:23 +0800
  • 1e6e909f88 [update] trim redundant comments jingyaogong 2026-09-20 20:33:40 +0800
  • b00be3e5ca
         Merge pull request #769 from qizwiz/fix/llm-response-unguarded jingyaogong 2026-09-20 20:19:39 +0800
  • 7f5ba40a4e
         Merge pull request #821 from lycDragonSlayer/refactor/remove-redundant-update-policy jingyaogong 2026-09-20 20:13:29 +0800
  • 949488fdb4
         Merge pull request #829 from DaoyuanLi2816/fix/dpo-paired-think-format jingyaogong 2026-09-20 20:07:46 +0800
  • e41bf7cacf
         Merge pull request #863 from Linxiushen/fix-fp16-attention-mask-nan jingyaogong 2026-09-20 19:49:22 +0800
  • b330a4d5c0
         Merge pull request #860 from Linxiushen/fix-tokenizer-pretrain-format jingyaogong 2026-09-20 19:47:52 +0800
  • b956bad49f
         Merge pull request #865 from Linxiushen/fix-requirements-missing-fastapi jingyaogong 2026-09-20 19:42:22 +0800
  • d698c06fc5
         Merge pull request #864 from Linxiushen/fix-nonstream-max-tokens jingyaogong 2026-09-20 19:42:05 +0800
  • 984a2c3611 [fix] requirements.txt 缺 fastapi/uvicorn，按 README 跑 serve_openai_api.py 直接崩 Linxiushen 2026-09-20 11:57:27 +0800
  • d4b67f1783 [fix] serve_openai_api 非流式端点的 max_tokens 不生效，与流式端点行为不一致 Linxiushen 2026-09-20 11:54:35 +0800
  • ba10c4d17b [fix] float16 下 attention mask 的 -1e9 溢出成 -inf，导致整条序列输出 NaN Linxiushen 2026-09-20 11:50:28 +0800
  • 7c2abda586 [feat] get_lr 支持可选的线性 warmup（默认关闭，行为不变） Linxiushen 2026-09-20 11:08:40 +0800
  • d6401cbb92 [fix] PretrainDataset 按 token id 屏蔽 padding，会误伤正文里字面的 <|endoftext|> Linxiushen 2026-09-20 10:36:32 +0800
  • cb08b9300b [fix] train_tokenizer 读预训练语料时产出 0 条文本，静默训出空词表 Linxiushen 2026-09-20 10:33:51 +0800
  • cc312c1cc6 [fix] unify Ctrl+C handling jingyaogong 2026-09-18 16:33:09 +0800
  • e2ee878519
         Merge pull request #857 from wbbeyourself/fix-ctrl-c-exception jingyaogong 2026-09-18 15:54:43 +0800
  • b63a9565e3 [fix] restore top-1 router gradient #858 jingyaogong 2026-09-18 15:28:50 +0800
  • 1aae25911b
         Merge pull request #859 from basil-k-aji-dev/fix/moe-router-gradient-topk1 jingyaogong 2026-09-18 15:03:36 +0800
  • 1c876435ef [fix] keep the MoE router trainable at num_experts_per_tok=1 basil-k-aji-dev 2026-09-17 21:31:41 +0530
  • 86472c9900
         Merge 856a57cacb into 7a9137d2e9 CharlesCheung996 2026-09-16 21:10:35 +0800
  • 856a57cacb docs: document Apple Silicon MPS training mengxin 2026-09-06 02:42:05 +0800
  • 791b5150cc test: cover MPS device and training paths mengxin 2026-09-06 02:41:54 +0800
  • 7b554a516b feat: enable MPS across training entrypoints mengxin 2026-09-06 02:41:42 +0800
  • 618b72dc2a feat: add device-agnostic training utilities mengxin 2026-09-06 02:41:29 +0800
  • 9fa620e2e5 fix: eval_llm.py Ctrl+C 退出异常修复 wbbeyourself 2026-09-16 17:38:56 +0800
  • c54a88887b
         Merge 6b2f4c939a into 7a9137d2e9 Justin Güse 2026-09-15 16:19:57 +0800
  • 7a9137d2e9 [update] trim verbose ddp comments jingyaogong 2026-09-15 14:53:11 +0800
  • 7988bc81a3
         Merge pull request #855 from basil-k-aji-dev/perf/ddp-followups jingyaogong 2026-09-15 14:51:37 +0800
  • 37ac206e9a [perf] skip the per-forward RoPE buffer broadcast in the RL trainers basil-k-aji-dev 2026-09-14 22:22:00 +0530
  • 21ec325dba [update] simplify safe math eval jingyaogong 2026-09-14 23:55:32 +0800
  • cdf0a93e02
         Merge pull request #847 from a0917-cell/fix/safe-math-eval jingyaogong 2026-09-14 23:34:28 +0800
  • f5b6d18c3d
         Merge 450885c0d5 into c367572702 Walle1999 2026-09-14 22:28:08 +0800
  • c367572702
         Merge pull request #854 from basil-k-aji-dev/fix/rl-ddp-grad-sync jingyaogong 2026-09-14 22:28:03 +0800
  • 9c088dd3a5 [fix] route RL trainer forward through DDP so gradients all-reduce basil-k-aji-dev 2026-09-14 14:54:07 +0530
  • a3c7b01cc0 [fix] keep agent rollout token-aligned #850 jingyaogong 2026-09-10 11:44:05 +0800
  • 655d4705eb feat: support Apple Silicon (MPS) for training and web demo Lingxu 2026-09-09 21:42:03 +0800
  • 59658fb60c fix(server): default `stream` to False to comply with OpenAI spec Lingxu 2026-09-09 21:27:48 +0800
  • 0f6803e551 [fix] replace tool eval() with an AST math evaluator tkgcc 2026-09-07 07:53:33 +0800
  • 9cb07dd1ec perf: enable fused AdamW in pretrain training neil.ge 2026-09-05 19:16:51 +0800
  • 859c8ee2a3 perf: enable fused AdamW in full SFT training neil.ge 2026-09-05 18:37:04 +0800
  • d8ff896a70
         Merge 864bd67437 into 6fc918beb6 hy 2026-09-04 22:30:24 +0800