Case Study: dapo-aime

The dapo-aime cookbook page demonstrates training math-reasoning ability on DAPO-Math-17k with direct GRPO (no SFT warm-start), using AIME 2024 as the primary benchmark and AIME 2025 / 2026 as auxiliary evals.

Recipe parameters


Algorithm	direct GRPO, no SFT warm-start
Base model	`Qwen/Qwen3-4B-Instruct-2507`
Training data	`BytedTsinghua-SIA/DAPO-Math-17k`, materialized locally to `data/train/full.jsonl`
Primary benchmark	AIME 2024 (`data/eval/aime2024.jsonl`)
Auxiliary evals	AIME 2025 / AIME 2026 (manifests provided on the same contract line as the primary benchmark)
Primary metrics	`eval_accuracy`, `eval_greedy_accuracy`, `eval_pass_at_k`

Why direct GRPO

Math problems have programmatically computable rewards — no reward model needed. GRPO's core idea is to sample the same prompt multiple times and normalize rewards within the group; this maps exactly to MinT's path from "reward / verifier / environment feedback" to the GRPO Recipe (loss_fn='importance_sampling').

Lifecycle order

uv sync → --dry-run → --eval-only → train

Same as chat-dpo: install dependencies first, dry-run the data path, run a baseline eval, then train for real. For fine-grained runnable commands, the cookbook points to the upstream README.

The corresponding approach on MinT

After you select the GRPO Recipe, MinT takes the loss_fn='importance_sampling' path. The panel's suggested hyperparameters are rank=32 / 5 epochs / lr=1e-4 / bs=8. A flow validated in the console can be migrated to a Python script.

Enterprise differences

Upgrade the base model to Qwen/Qwen3-30B-A3B-Instruct-2507 or Qwen/Qwen3-4B-Thinking-2507 (reasoning / chain-of-thought variant, see Supported Models).
For MoE models, a full fine-tune averages about 1 to 1.5 hours (30B model, A3B structure).
Save a checkpoint after training: each checkpoint carries metadata (Recipe, data version, hyperparameters, eval scores) for easy traceability.

For the RL and GRPO algorithms themselves, see RLHF / GRPO; for a DPO example, see chat-dpo.