tiny-recursive-model-sudoku-mvp27

MVP27 Sudoku solver adapters (LoRA + explicit heads) trained on top of Qwen3-8B.

This repository contains three checkpoints organized by experiment family:

• mvp27-7065-raw-baseline-ckpt2400
• mvp27-7621-closure-continuation-ckpt800
• mvp27-7618-proj-heavy-ckpt400

Contents in each subfolder

Each subfolder typically includes:

• adapter_model.bin
• adapter_config.json
• explicit_heads.pt
• tokenizer files (for reproducibility)

Base model

All adapters are intended for:

• Base model: Qwen/Qwen3-8B

Quick loading example (PEFT)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "Qwen/Qwen3-8B"
adapter_subdir = "mvp27-7065-raw-baseline-ckpt2400"
repo_id = "sf655656/tiny-recursive-model-sudoku-mvp27"

tok = AutoTokenizer.from_pretrained(base_model_id, use_fast=True)
base = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
)
model = PeftModel.from_pretrained(base, repo_id, subfolder=adapter_subdir)
model.eval()

Notes

• These are research checkpoints for Sudoku-specific experiments.
• Some local training artifacts are intentionally excluded from upload.
• Evaluation scripts and logs live in the training project workspace.

Contact

If you need exact run configs, metrics, or reproducibility details, please open an issue in the project workspace/repo that produced these checkpoints.