← back to report

jingyaogong/minimind

// all events — 2026-09-28

Type AI Score Description Actor Reason Date
commit 1.00 [fix] replace tool eval() with an AST math evaluator Commit message contains explicit AI assi 2026-09-06
commit 1.00 [fix] PPO rollout mode Commit message contains explicit AI assi 2026-07-23
pr 1.00 [feat] get_lr 增加可选 warmup_frac 参数(默认关闭,行为逐值不变) Explicitly states that the PR was comple 2026-09-20
pr 1.00 [fix] float16 下 attention mask 的 -1e9 溢出成 -inf,导致整条序列输出 NaN Explicitly states the PR was completed w 2026-09-20
pr 1.00 [fix] train_tokenizer 读预训练语料时产出 0 条文本,静默训出空词表 Explicitly states the PR was completed w 2026-09-20
pr 1.00 [feat] get_lr 支持可选的线性 warmup(默认关闭,行为逐值不变) Explicitly states the PR was completed w 2026-09-20
pr 1.00 [fix] requirements.txt 缺 fastapi/uvicorn,按 README 跑 serve_op Explicitly states the PR was completed w 2026-09-20
pr 1.00 [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 Explicitly states the PR was completed w 2026-09-20
pr 1.00 [fix] PretrainDataset 按 token id 屏蔽 padding,会误伤正文里字面的 <|endo Explicitly states the PR was completed w 2026-09-20
pr 1.00 feat: add complete Apple Silicon MPS training support 明确包含 GPT-6 的 Co-Authored-By AI 协作声明 2026-09-05
pr 1.00 [fix] 用 AST 求值器替换工具里的 eval(),并修复 Windows 上工具调用全部静默失败 Explicitly states it was generated with 2026-09-06
pr 1.00 fix: don't char-slice the prompt by max_new_tokens / max_tok PR body explicitly mentions AI collabora 2026-06-16
pr 0.95 [refactor] remove redundant rollout update_policy after torc The content is explicitly presented as a 2026-08-24
pr 0.75 Codex searchshortqa deepspeed The Codex label strongly suggests an AI- 2026-06-02
pr 0.75 Codex/code agent rl The explicit Codex/code-agent wording st 2026-07-12
pr 0.28 docs: 完善中英双语数据集指南 / expand bilingual dataset guide Polished bilingual summary is somewhat b 2026-09-01
pr 0.25 MoE dispatch: one host sync per layer instead of 3×num_exper Detailed benchmarks, implementation refe 2026-09-03
pr 0.20 Expose num_experts / num_experts_per_tok / seed on the train Polished and structured, but includes co 2026-09-03
pr 0.20 [docs] add beginner-friendly learning tutorial (docs/) Domain-specific detail, informal tone, n 2026-07-31
pr 0.20 docs: improve English README framing for sharing and onboard Uses domain-relevant phrasing, clear pur 2026-05-23
pr 0.18 Add CE regularization term to DPO loss Technical Chinese explanation and concre 2026-08-12
pr 0.15 [fix] Complete streaming response fields and stop after gene Polished structure, but it remains concr 2026-09-22
pr 0.15 chore(deps): 使用uv pyproject.toml替代requirements.txt管理依赖;增加vsc Specific dependency and tooling changes, 2026-05-05
pr 0.15 docs: state the measurement resolution of the objective benc Detailed reproduction notes and candid u 2026-09-03
pr 0.12 Harden calculate_math tool handler against code injection (C Highly specific security details and cod 2026-05-11
commit 0.10 [fix] repetition_penalty Message is terse, uses domain-specific l 2026-05-07
commit 0.10 [update] params log Bare-bones commit messages typical of hu 2026-01-07
commit 0.10 [update] mask log Bare-bones commit messages typical of hu 2026-01-07
commit 0.10 [update] readme Bare-bones commit messages typical of hu 2026-01-06
commit 0.10 [update] simplify loader Bare-bones commit messages typical of hu 2026-01-05
commit 0.10 [update] readme Bare-bones commit messages typical of hu 2026-01-05
commit 0.10 [update] rename train tokenizer Bare-bones commit messages typical of hu 2026-01-05
commit 0.10 [update] readme Bare-bones commit messages typical of hu 2026-01-05
pr 0.10 [fix] keep DPO preference-pair think formatting aligned Detailed domain-specific reasoning and v 2026-08-30
pr 0.10 feat: 添加 macOS Metal (MPS) 设备支持 Contains technical jargon, code referenc 2026-05-25
pr 0.10 add train_agentic_rl_async.py 一个是 MiniMind 在多轮 Tool-Use 场景 Contains project-specific Chinese techni 2026-05-28
pr 0.10 [fix] 更新.gitignore文件,添加忽略目录;修改web_demo.py,优化样式和功能 Brief, terse, uses domain-specific terms 2026-05-10
pr 0.10 一些小的改动 Succinct, direct, includes domain-specif 2026-04-12
pr 0.08 [fix] unpack MoE expert weights when converting back to PyTo Detailed, domain-specific Chinese descri 2026-09-23
pr 0.08 [fix] save final checkpoints after the residual accumulation Thorough technical report with specific 2026-09-22
pr 0.08 Add MiniMind source course materials A concise, repository-specific title wit 2026-06-28
pr 0.08 fix(server): default `stream` to False to comply with OpenAI Specific API behavior, reproduction step 2026-09-09
pr 0.07 [fix] align LoRA inference paths with training output Specific repository terminology and conc 2026-09-22
pr 0.06 fix: 修复多语言重复惩罚 / make repetition penalty multilingual Bilingual technical summary uses reposit 2026-09-01
commit 0.05 [update] readme content Terse update with no noticeable AI-style 2026-08-21
commit 0.05 [update] readme content Terse update with no noticeable AI-style 2026-08-21
commit 0.05 [update] readme content Terse update with no noticeable AI-style 2026-08-21
commit 0.04 [docs] 在客观评测表下补充各数据集的测量分辨率 Concise Chinese technical summary with c 2026-09-03
pr 0.04 [fix] route RL trainer forward through DDP so gradients all- Highly specific reproduction and impleme 2026-09-14
pr 0.04 docs: 添加多语言 README 入口 / add multilingual README entry points Specific bilingual technical details and 2026-09-01
commit 0.03 [fix] residual step order Terse, domain-specific commit message wi 2026-09-22
commit 0.03 [perf] skip the per-forward RoPE buffer broadcast in the RL Specific DDP details and test results in 2026-09-14
commit 0.03 [fix] route RL trainer forward through DDP so gradients all- Specific DDP details and informal techni 2026-09-14
commit 0.03 [update] 把随机种子暴露到命令行 Specific implementation references and t 2026-09-03
commit 0.03 [update] 把 num_experts / num_experts_per_tok 暴露到命令行 Concrete entry points and defaults make 2026-09-03
pr 0.03 fix: guard against empty/filtered LLM responses in chat_api. Polished but includes concrete provider 2026-05-18
pr 0.03 [perf] skip the per-forward RoPE buffer broadcast in the RL Specific technical discussion, tests, an 2026-09-14
pr 0.03 [fix] keep the MoE router trainable at num_experts_per_tok=1 Specific technical diagnosis, measuremen 2026-09-17
pr 0.03 feat: support Apple Silicon (MPS) for training and web demo Specific implementation details, benchma 2026-09-09
pr 0.03 perf: enable fused AdamW in SFT & pretrain training 内容简短具体,含 CUDA、AdamW 和回退路径等领域细节。 2026-09-05
pr 0.03 [feat] add MPS support for Apple Silicon Specific benchmarks and implementation d 2026-08-13
commit 0.02 fix: update residual gradients before saving final checkpoin Concise technical commit message; no sig 2026-09-22
commit 0.02 fix: align LoRA inference paths with training output Short domain-specific commit message wit 2026-09-22
commit 0.02 [fix] keep the MoE router trainable at num_experts_per_tok=1 Detailed domain-specific explanation wit 2026-09-17
commit 0.02 [update] trim verbose ddp comments Terse, natural commit message with a spe 2026-09-15
commit 0.02 [update] simplify safe math eval Terse, issue-focused commit message with 2026-09-14
commit 0.02 [update] remove expert args Terse, issue-specific commit message wit 2026-09-04
commit 0.02 fix: guard against empty/filtered LLM responses in chat_api. Concise technical description with speci 2026-05-18
pr 0.02 chore: remove .idea and add to .gitignore Very terse conventional chore title with 2026-08-27
commit 0.01 [update] trim redundant comments Terse commit-style wording with no AI-li 2026-09-20
commit 0.01 [fix] requirements.txt 缺 fastapi/uvicorn,按 README 跑 serve_op Specific reproduction details, file refe 2026-09-20
commit 0.01 [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 Detailed domain-specific diagnosis and t 2026-09-20
commit 0.01 [fix] float16 下 attention mask 的 -1e9 溢出成 -inf,导致整条序列输出 NaN Concrete dtype behavior, reproduction co 2026-09-20
commit 0.01 [fix] train_tokenizer 读预训练语料时产出 0 条文本,静默训出空词表 Specific dataset formats, counts, and va 2026-09-20
commit 0.01 [fix] keep agent rollout token-aligned #850 Terse, domain-specific title with an iss 2026-09-10
commit 0.01 [fix] keep DPO preference pair think formatting aligned Brief informal commit message with proje 2026-08-30
commit 0.01 [refactor] remove redundant rollout update_policy after torc Concise refactor summary using repositor 2026-08-24
commit 0.00 [fix] unify Ctrl+C handling Terse, specific fix message with no AI-s 2026-09-18
commit 0.00 [fix] restore top-1 router gradient #858 Concise technical commit title referenci 2026-09-18
commit 0.00 fix: eval_llm.py Ctrl+C 退出异常修复 Brief Chinese fix message with a concret 2026-09-16
commit 0.00 [update] readme — 2026-08-31
commit 0.00 [update] dependency version — 2026-08-31
commit 0.00 [update] dependency version — 2026-08-29
commit 0.00 [update] supporter charts Terse, informal commit message typical o 2026-08-06
commit 0.00 [refactor] PPO cleanup Terse commit message, domain-specific ab 2026-07-23
commit 0.00 [fix] PPO: align mb_resp_logp with old_resp_logp under autoc Technical detail, domain jargon, informa 2026-07-21
commit 0.00 [fix] robustness Very terse commit, no sign of AI phrasin 2026-06-28
commit 0.00 [fix] ddp exit hang Terse commit message with shorthand and 2026-06-01
commit 0.00 [fix] lora dims — 2026-05-31
commit 0.00 [fix] issue#771 Very terse commit message; no AI signs p 2026-05-19
commit 0.00 [update] requirements Terse, typical human commit message with 2026-05-16
commit 0.00 [feat] reduce RL memory Very terse commit message typical of hum 2026-05-06
commit 0.00 fix: resolve CUDA OOM in train_grpo.py on GPUs with <=8GB VR Contains domain-specific jargon and abbr 2026-05-06
commit 0.00 [update] readme Generic commit subject, no stylistic AI 2026-05-06
commit 0.00 [fix] inference bug Terse, domain-specific commit message; n 2026-05-03
commit 0.00 [update] readme Very terse message; likely human-written 2026-05-01
commit 0.00 [update] readme Very terse, typical human commit message 2026-04-28
commit 0.00 [fix] bugs Message is terse and uses standard commi 2026-04-27
commit 0.00 [fix] rollout bugs Short, domain-specific, and informal com 2026-04-27
commit 0.00 [update] readme Brief and standard commit note, no AI ha 2026-04-27
commit 0.00 [update] readme Very terse; typical human commit message 2026-04-24
commit 0.00 [update] readme Short, domain-specific; no AI hallmarks 2026-04-24
commit 0.00 [update] readme Commit messages are terse and typical of 2026-04-21
commit 0.00 [fix] bugs Short, domain-typical, and lacking AI-st 2026-04-21
commit 0.00 [fix] transformers-5.x Very terse commit, standard human brevit 2026-04-19
commit 0.00 [update] tie embedding Brief domain abbreviation, no AI stylist 2026-04-19
commit 0.00 [update] readme Standard minimalist commit, lacks AI phr 2026-04-19
commit 0.00 [update] readme Very terse commit message; no AI signals 2026-04-10
commit 0.00 [fix] lora compile All items are terse, domain-specific, an 2026-04-09
commit 0.00 [update] readme Minimal, template-style message with no 2026-04-09
commit 0.00 [update] readme Short and standard commit message, typic 2026-04-09
commit 0.00 [update] bench Concise project-specific update, no AI s 2026-04-09
commit 0.00 [fix] lora moe Direct issue fix; brief and domain-relev 2026-04-09
commit 0.00 [update] add a comment Human-like succinct edit description wit 2026-04-09
commit 0.00 [update] image Terse commit with no AI signals, likely 2026-04-04
commit 0.00 [update] readme Brief, typical human commit message stru 2026-04-04
commit 0.00 [update] open causal Very terse, typical human commit style; 2026-04-02
commit 0.00 [update] readme Extremely brief, standard for human comm 2026-04-02
commit 0.00 [update] readme Extremely terse and typical of human com 2026-04-01
commit 0.00 [update] default model Terse and minimal commit message, typica 2026-03-31
commit 0.00 [update] random seed Very terse, standard commit message styl 2026-03-27
commit 0.00 [update] fp16 inference Brief, domain-specific commit message. 2026-03-27
commit 0.00 [update] image Terse, mechanical commit; typical human 2026-03-26
commit 0.00 [update] change default seq_len Concise and technical; common human comm 2026-03-26
commit 0.00 [update] minimind-3 Very brief, informal; no AI hallmarks pr 2026-03-24
commit 0.00 [fix] align log/save last-step check and ETA with 1-indexed Specific technical fix; human-like jargo 2026-03-24
commit 0.00 [fix] gradient accumulation step alignment Direct technical fix; human engineering 2026-03-24
commit 0.00 [update] empty_think_ratio Minimal, to-the-point; lacks AI politene 2026-02-06
commit 0.00 [update] empty_think_ratio Repetitive, brief update; likely human c 2026-02-05
commit 0.00 [feat] data process Ambiguous but informal; typical human sh 2026-02-05
commit 0.00 [update] save interval Concise and clear; common human commit m 2026-01-30
commit 0.00 [update] safe half Very brief, technical term; no AI stylis 2026-01-30
commit 0.00 [fix] data skip Extremely terse and informal phrasing ty 2026-01-18
commit 0.00 [update] shuffle data Brief, informal update note with common 2026-01-18
commit 0.00 [fix] max length Minimalist technical fix; lacks any poli 2026-01-17
commit 0.00 [update] pretrain load Concise, technical update note with stan 2026-01-17
commit 0.00 [update] align mask Very short and specific; common ML/engin 2026-01-15
commit 0.00 [update] align loss Terse technical update; no hallmark AI p 2026-01-14
commit 0.00 [fix] compile unpack Brief human-style fix message ('fix comp 2026-01-14
commit 0.00 [feat] add compile Succinct feature addition; common Git co 2026-01-14
commit 0.00 [update] prompt prefill Direct technical update; uses specific j 2026-01-13
commit 0.00 [update] show speed Short, informal update typical of human 2026-01-07
commit 0.00 [update] rename reason Bare-bones commit messages typical of hu 2026-01-05
pr 0.00 [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.tom — 2026-09-26
pr 0.00 [update] 新增一键训练 pipeline 脚本 run_pipeline.sh — 2026-09-26
pr 0.00 [update] 新增一键训练 pipeline 脚本 run_pipeline.sh — 2026-09-26
pr 0.00 [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.tom — 2026-09-26
pr 0.00 Fix readme windows details — 2026-06-01
pr 0.00 feat: add multi-turn tool-use on-policy distillation — 2026-09-23
pr 0.00 fix: eval_llm.py Ctrl+C 退出异常修复 Concise Chinese fix title naming the aff 2026-09-16
pr 0.00 [fix] PPO: align mb_resp_logp with old_resp_logp under autoc In-depth technical Chinese, domain detai 2026-07-21
pr 0.00 [fix] PPO: align mb_resp_logp with old_resp_logp under autoc In-depth technical Chinese, domain detai 2026-07-21
pr 0.00 [fix] PPO: align mb_resp_logp with old_resp_logp under autoc In-depth technical Chinese, domain detai 2026-07-21
pr 0.00 feat: 增加通用 On-Policy Distillation(GKD / OPD)训练器 Longform technical Chinese, accurate dom 2026-07-15
pr 0.00 Add on-device QAT training script, optimize the training rec Contains jargon, informal tone, and deta 2026-06-30
pr 0.00 取消模型上下文限制,增加模型动态长度扩展机制,并保持前向兼容性 Detailed technical feature write-up, sty 2025-07-09
pr 0.00 add learning notes Generic commit title; lacks AI-like phra 2026-06-02
pr 0.00 use wandb — 2026-05-29
pr 0.00 适配昇腾NPU Text shows informal tone, domain referen 2026-05-18
pr 0.00 fix: resolve CUDA OOM in train_grpo.py on GPUs with <=8GB VR Filled-out template with specific techni 2026-05-06
pr 0.00 Main No content to analyze; title only with n 2026-05-02
pr 0.00 使用einops进一步提升代码可读性 Title is concise, domain-specific, and i 2026-03-27
pr 0.00 Wz test Extremely terse; no AI hallmarks, resemb 2026-04-05
pr 0.00 [feat] add dapo algorithm Detailed technical phrasing, human tone, 2026-03-27
pr 0.00 merge redundant forward passes for logps and aux_loss (in tr — 2026-03-24
pr 0.00 添加数据集加载逻辑、网页内容抓取与数据预处理逻辑 — 2025-12-16
pr 0.00 [docs] Fix wording in RLHF section of README.md file — 2026-01-27
pr 0.00 [docs]: clarify pretraining data format in README — 2025-12-31
pr 0.00 refactor: optimize tensor wrapping in lm_dataset.py — 2025-12-20
pr 0.00 [fix] 修复训练脚本中 1-indexed step 与 0-indexed 逻辑混用的问题 — 2026-03-24
pr 0.00 Fix SFT resume with torch.compile enabledfix sft resume with — 2026-03-19
pr 0.00 新建test分支 — 2026-03-22
pr 0.00 [mod & add] fix spo algorithm, add dapo and cispo algorithm — 2026-01-30
pr 0.00 Mega — 2025-06-26
pr 0.00 Add dynamic growth pipeline, eval tooling, and overnight run — 2026-02-22
pr 0.00 更新了 model / trainer 中的注释 & PyTorch 新版本 Automatic Mixed Preci — 2026-02-04
pr 0.00 Update requirements.txt — 2024-10-28
pr 0.00 Auto tokenizer name path fix — 2024-10-03
pr 0.00 Update requirements — 2024-10-01
pr 0.00 [add] add gating term on po algorithm — 2026-02-03
pr 0.00 add muon optimizer — 2026-02-02
pr 0.00 modified: .gitignore — 2026-01-19
pr 0.00 Fix DPO loss_mask boundary (include first assistant token) — 2026-01-07
pr 0.00 [feat] Support Minimind retrieval-augmented generation (RAG) — 2025-11-16
pr 0.00 perf: merge LoRA weights into model for inference — 2025-12-23
pr 0.00 Create RESOURCES.md for MiniMind project — 2025-10-24
pr 0.00 Perf/merge lora weights — 2025-12-23
pr 0.00 Fix: support loading DDP-saved LoRA weights for inference — 2025-12-22
pr 0.00 [feat] add interactive notebook — 2025-02-23
pr 0.00 [feat] Add Training Web UI — 2025-11-06
pr 0.00 fix: 调整model_lora.py里面lora作用的对象 — 2025-12-05
pr 0.00 Add attention gate — 2025-12-12
pr 0.00 feat: 增加 LoRA alpha 缩放系数及命令行支持 — 2025-12-12
pr 0.00 [fix] Refactor get_lr function to include min_lr calculation — 2025-12-06
pr 0.00 feat: add merge_lora.py to support merging LoRA weights into — 2025-12-05
pr 0.00 第一次尝试 — 2025-11-21
pr 0.00 fix: attn_forwad when is_causal=True assert attn_mask is Non — 2025-11-18
pr 0.00 Train_Grpo 添加注释 — 2025-10-26
pr 0.00 [feat] update model install method — 2025-11-05
pr 0.00 [feat] add MNN support to README. — 2025-11-10