Small models, post-trained deliberately.
A Go reproduction of the VibeThinker pipeline: the Spectrum-to-Signal recipe that takes small dense reasoning models through diversity-exploring SFT, expert fusion, and MaxEnt-guided reinforcement learning.
Research reproduction. Every stage of the published method is implemented on the mlx-go stack with property tests over the design's correctness invariants, and the full recipe runs end to end on toy configurations. Real-model evaluation requires weights you provide; none are redistributed. The canonical real-model result so far is parity with the base model, not an improvement.
What it does- Spectrum SFT — diversity-exploring supervised fine-tuning that probes checkpoints per subdomain with an unbiased Pass@K estimator.
- Expert fusion — merges subdomain-specialist checkpoints into one SFT model via weighted parameter averaging over safetensors.
- MaxEnt-guided RL — modulates GRPO advantages with a max-entropy-deviation weight, focusing optimization on prompts with high exploratory value.
- Offline self-distillation — filters correct reasoning traces by length-normalized likelihood to build the next round of student data.
- Evaluation harness — commands for method comparison, toy end-to-end runs, and real-model evaluation against user-provided weights.
This port tests whether a published post-training recipe can be reproduced on mlx-go and the mlx-go-lm model registry. External services stay behind interfaces, and the evaluation reports parity when no improvement is measured. mlx-go-sia applies a separate improvement loop to agents; mlx-go-fedistill studies training across peers.