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

From run 012 on, each seed run is one directory, runs/<experiment>_s<seed>; where present, runs/<experiment> holds only the merged multi-seed summary. Earlier runs are single-seed directories. The autoresearch loop uploads every seed run when it finishes. previous_attempts/ keeps an archived, cancelled attempt whose rerun is the scored one, and pretrain_poc/ the pretraining proof of concept. The earliest runs predate stage checkpoints and have only final.pt, or no weights at all. 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.

Datasets

data/<name>/ holds gzipped JSONL datasets from the nd-rl repository. The code that built each one is in that repository under code/experiments/current/<name>/, and docs/DATASETS.md there describes them.

  • data/long-proofs-20261013/: natural-deduction theorems over up to 10 atoms with 3-10 premises, certified to need at least 20 body lines (lines after the premise block), plus theorems with exact minimal lengths 2-19 (*_f20.jsonl, tight.jsonl), and approach E's 35-80-line proofs with a signature lower bound (e_raw.jsonl, e_raw_sig.jsonl).
  • data/long-rl-20261014/: the long-RL experiment's data. pretrain.jsonl.gz is Dan Pandori's cap-6 proofs (from dan-pandori/nd-takehome at 51604b3) with distractor premises; rl/dev/holdout are the RL target, dev and held-out pools, built from the long-proofs set and, for rl, Dan's ladder RL targets. Every proof was checked by Lean.

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. Checkpoints do not embed a full architecture configuration. Most run directories include source/ with the exact pretrain.py, harness.py, and remote.sh that produced them. For the earliest runs, the top-level 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. Do not assume every run uses the same architecture. The external training fork is dan-pandori/nd-takehome, pinned in each run's source/remote.sh (FORK_COMMIT): ab629c7d2cdf3dc50d471e302d14643ba0782749 for runs before 098, and 51604b3feaa5e66821d1faba22238fe6d96d46d0 (Lean 4 tokenizer and reward) from 098 on.

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.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading