---

name: minimind description: "Routes MiniMind LLM architecture, tokenizer, training, inference, serving, model conversion, RLHF, RLAIF, and Agentic RL workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MiniMind Repo Skill

Use this skill for MiniMind tasks: training a tiny language model from scratch, validating MiniMind JSONL data, using the MiniMind tokenizer/chat template, running inference, serving an OpenAI-compatible local API, converting artifacts for third-party engines, or planning DPO/PPO/GRPO/CISPO/Agentic RL post-training.

MiniMind is an educational and practical PyTorch LLM codebase for small Dense and MoE decoder-only models aligned with Qwen3/Qwen3-MoE conventions. It includes custom tokenizer assets, pretrain/SFT/LoRA/RL training entrypoints, tool-call and thinking chat-template support, OpenAI-compatible serving helpers, and conversion workflows.

Before you choose a route

1. Identify whether the user is working with data, raw PyTorch weights, Transformers-format model directories, training scripts, post-training objectives, or serving/API clients.
2. Check whether real model weights, datasets, reward-model checkpoints, or external serving engines already exist locally. This skill does not assume downloads are available.
3. Treat CUDA as the practical backend for real training and meaningful generation. CPU is useful for imports, validators, parser checks, and tiny random-tensor smokes.
4. Use bundled validators and smoke helpers before long training, server, conversion, or RL runs.

Sub-skill routes

• training-basics: use for tokenizer compatibility, pretrain JSONL, SFT/LoRA conversation JSONL, MiniMindConfig, core pretraining, full SFT, LoRA training, checkpoint resume, DDP, output weight naming, and tiny model smokes.
• inference-serving: use for local generation, raw .pth versus Transformers artifact decisions, LoRA inference/merge/export, OpenAI-compatible /v1/chat/completions, Streamlit-style UI behavior, tool-call parsing, thinking output, model conversion, and vLLM/llama.cpp/Ollama-style serving.
• rlhf-agentic: use for white-box distillation, DPO, PPO, GRPO/CISPO, Agentic RL, RLAIF/Agentic JSONL schemas, reward models, rollout engines, SGLang planning, tool-use rewards, and post-training weight handoff.

Shared references and helper

• Read references/package-overview.md for the repository capability map, artifact naming, dependencies, and evidence-backed workflow summary.
• Read references/troubleshooting.md for cross-cutting install/import/backend/data/artifact triage before drilling into a sub-skill's troubleshooting page.
• Read references/repo-provenance.md before deciding whether this skill is stale for a checkout.
• Run scripts/check_minimind_environment.py to inspect core dependency availability, CUDA, Qwen3 class support, and optional MiniMind source-module imports without loading large weights.

Minimal public setup checks

For a general MiniMind Python environment, install dependencies from the repository's requirement list and a backend-appropriate PyTorch build. The requirement list intentionally leaves torch commented because the right wheel depends on CPU/CUDA/ROCm and host drivers. The OpenAI-compatible server additionally needs fastapi and uvicorn.

After installation, run a lightweight dependency/backend check:

python scripts/check_minimind_environment.py --check-qwen3 --check-cuda

If you are working inside a local MiniMind checkout or an unpacked source tree and want module import checks, pass that tree explicitly:

python scripts/check_minimind_environment.py --module-root MINIMIND_SOURCE_TREE --check-modules

Do not treat this import check as proof that a long training job, model download, reward-model load, or third-party serving engine will work. Route to the relevant sub-skill for the workflow-specific validation.

Common route decisions

User goal	Route	First safe action
Validate pretrain_t2t_mini.jsonl, sft_t2t_mini.jsonl, or custom LoRA data	training-basics	Run the bundled JSONL validator for the expected schema.
Start or resume pretraining/SFT/LoRA	training-basics	Validate data, run a tiny smoke, confirm input/output weight prefixes and backend.
Chat with a trained MiniMind model	inference-serving	Classify the artifact as raw .pth or Transformers format.
Serve MiniMind behind an OpenAI-compatible endpoint	inference-serving	Prefer a Transformers directory, check API dependencies, then probe one request.
Debug <tool_call> or <think> parsing	inference-serving	Run the bundled tool-call smoke helper on representative text.
Convert raw weights for vLLM, llama.cpp, or Ollama	inference-serving	Validate artifacts and plan raw-to-Transformers or Qwen3-compatible export first.
Run DPO, PPO, GRPO/CISPO, or Agentic RL	rlhf-agentic	Validate post-training JSONL schema and external reward/teacher/checkpoint prerequisites.
Use optional SGLang rollout for RL	rlhf-agentic	Confirm the local SGLang service, logprob support, update endpoint, and dedicated shared checkpoint path; otherwise use torch rollout.

Backend and dependency notes

• CUDA is a required practical backend for the selected training/inference capability surface; tiny CPU checks only validate plumbing.
• The native evidence supports dense and MoE MiniMind configs; checkpoint names add _moe for MoE variants.
• MiniMind tokenizer behavior depends on tokenizer_config.json with chat-template support for roles, <think>, <tool_call>, and <tool_response>.
• FastAPI/Uvicorn are serving dependencies; Streamlit is a UI dependency; SGLang/vLLM/llama.cpp/Ollama are optional external engines and are not part of the minimum environment.

Freshness check

Before applying this skill to a changed MiniMind checkout, read references/repo-provenance.md. If the commit, dirty state, tokenizer/model files, training scripts, or public workflow docs differ materially from the recorded snapshot, refresh this repo skill instead of assuming the old guidance is current.