robbiethompson2018/nd-rl-checkpoints

Natural deduction autoresearch checkpoints

Experimental small proof models from the nd-rl pretraining autoresearch workstream. This repository shares checkpoint weights and their recorded experiment summaries.

Checkpoint stages

• runs/<run>/final.pt: after pretraining and reinforcement learning, with final evaluation completed. These are not pretraining-only weights.
• Future pretrain.pt files: immediately after pretraining, before RL.
• Future rl_round_XX.pt files: after the named RL round.

The initial upload contains historical final checkpoints only. Each summary.json records the run's status, source hashes, budgets, and evaluation results. Scores across different harness hashes or budgets should not be treated as directly comparable.

Download and load

hf download robbiethompson2018/nd-rl-checkpoints --include 'runs/001_gpu-batching/*' --local-dir checkpoints

import torch

checkpoint = torch.load(
    'checkpoints/runs/001_gpu-batching/final.pt',
    map_location='cpu',
    weights_only=True,
)
# Construct the matching model architecture, then:
# model.load_state_dict(checkpoint['state'])
# checkpoint['tok_mode'] specifies the tokenizer mode.

These are raw PyTorch state dictionaries, not Transformers from_pretrained packages. They do not contain optimizer or RNG state for exact training resumption. Historical files do not embed a full architecture configuration. The sources/ folder contains available pretraining implementations; match the first 12 hexadecimal characters of a file's SHA-256 to pretrain_sha in the run summary. baseline.py matches run 001_gpu-batching, 010_replicate-seed1, and the s20_baseline_* runs. For implementations not included here, obtain the matching code from the project maintainers. Do not assume every run uses the same architecture. The external training fork is dan-pandori/nd-takehome, pinned to ab629c7d2cdf3dc50d471e302d14643ba0782749 for these runs.

The autoresearch metric uses development theorems and is not a final held-out benchmark. No inference service or hosted GPU is provisioned by this repository.