How jingyaogong/minimind works
MiniMind is a from-scratch, compact language-model project: its main user journey is preparing training data, training or adapting a model, then evaluating or serving generated conversations. The repository also provides distinct preference/RL, distillation, tool-use, tokenizer, and chat-interface capabilities. The graph preserves these documented branches while limiting connections to relationships visible in excerpts, the source index, or README; unsampled implementation details are not inferred.
Main entry points: model_minimind.py Dense and MoE model · lm_dataset.py Dataset loaders · rollout_engine.py Rollout engines · model_lora.py LoRA adapters
Components 23
Data and tokenization
- Training corpora
dataset/dataset.md - Tokenizer training
trainer/train_tokenizer.py - Dataset loaders
dataset/lm_dataset.py
Model training
- Pretraining
trainer/train_pretrain.py - SFT and LoRA
trainer/train_full_sft.py - LoRA training
trainer/train_lora.py - DPO training
trainer/train_dpo.py - PPO and GRPO
trainer/train_ppo.py - GRPO training
trainer/train_grpo.py - Agent RL
trainer/train_agent.py - Distillation
trainer/train_distillation.py - Rollout engines
trainer/rollout_engine.py - Training utilities
trainer/trainer_utils.py
Model runtime
- Dense and MoE model
model/model_minimind.py - LoRA adapters
model/model_lora.py - Weight conversion
scripts/convert_model.py
Evaluation and chat
- Model evaluation
eval_llm.py - Tool-call evaluation
scripts/eval_toolcall.py - OpenAI API server
scripts/serve_openai_api.py - Chat Web UI
scripts/web_demo.py - Chat API example
scripts/chat_api.py
Other components
- Engineer or user
- SGLang service
Connections 36
- Engineer or user → Tokenizer training: trains tokenizer
README.md - Engineer or user → Pretraining: starts training
README.md - Engineer or user → Model evaluation: enters prompt
eval_llm.py - Engineer or user → OpenAI API server: sends requests
README.md - Engineer or user → Chat Web UI: chats
README.md - Training corpora → Dataset loaders: provides examples
dataset/lm_dataset.py - Pretraining → Dataset loaders: loads data
trainer/train_pretrain.py - SFT and LoRA → Dataset loaders: loads data
trainer/train_full_sft.py - LoRA training → Dataset loaders: loads data
trainer/train_lora.py - DPO training → Dataset loaders: loads preferences
trainer/train_dpo.py - PPO and GRPO → Dataset loaders: loads data
trainer/train_ppo.py - GRPO training → Dataset loaders: loads data
trainer/train_grpo.py - Agent RL → Dataset loaders: loads data
trainer/train_agent.py - Distillation → Dataset loaders: loads data
trainer/train_distillation.py - Pretraining → Dense and MoE model: trains
trainer/train_pretrain.py - SFT and LoRA → Dense and MoE model: trains
trainer/train_full_sft.py - LoRA training → LoRA adapters: applies adapters
trainer/train_lora.py - DPO training → Dense and MoE model: trains policy
trainer/train_dpo.py - PPO and GRPO → Dense and MoE model: trains actor
trainer/train_ppo.py - GRPO training → Dense and MoE model: trains policy
trainer/train_grpo.py - Agent RL → Dense and MoE model: trains agent
trainer/train_agent.py - Distillation → Dense and MoE model: trains student
trainer/train_distillation.py - Agent RL → Rollout engines: uses engine
trainer/train_agent.py - PPO and GRPO → Rollout engines: uses engine
trainer/train_ppo.py - GRPO training → Rollout engines: uses engine
trainer/train_grpo.py - Rollout engines → SGLang service: sends HTTP requests
trainer/rollout_engine.py - Training utilities → Dense and MoE model: initializes model
trainer/trainer_utils.py - Model evaluation → Dense and MoE model: loads and generates
eval_llm.py - Model evaluation → LoRA adapters: loads adapters
eval_llm.py - Tool-call evaluation → Dense and MoE model: evaluates calls
scripts/eval_toolcall.py - OpenAI API server → Dense and MoE model: serves inference
scripts/serve_openai_api.py - OpenAI API server → LoRA adapters: supports adapters
scripts/serve_openai_api.py - Chat Web UI → Training utilities: uses utilities
scripts/web_demo.py - Weight conversion → Dense and MoE model: loads base weights
scripts/convert_model.py - Weight conversion → LoRA adapters: merges adapters
scripts/convert_model.py - Tokenizer training → Engineer or user: produces tokenizer
README.md