G-NRI: Symmetry-Restored Neural Rule Induction
Reference checkpoints for the G-NRI public release. The repository ships the
eight gnri-condition training seed checkpoints (seeds 42, 43, 44, 45, 46, 47, 48, 49, one
checkpoint_last.pt each), a SHA256 manifest (manifest_sha256.json), and
inference-only weights (model.safetensors + config.json) loadable with
RuleInducer.from_pretrained.
Usage: inference-only weights
import torch
from gnri import RuleInducer
model = RuleInducer.from_pretrained("phuayj/gnri")
model.eval()
Usage: full training checkpoints
from gnri import load_checkpoint_from_hub
# Seeds are stored as s42/... s49/ under checkpoint_last.pt
model = load_checkpoint_from_hub("phuayj/gnri", filename="s42/checkpoint_last.pt")
Or download the file directly:
from huggingface_hub import hf_hub_download
path = hf_hub_download("phuayj/gnri", "s42/checkpoint_last.pt")
The full checkpoints can be passed to the release evaluation entry points, e.g.
python evaluate_uci.py --checkpoint <path> --condition eqarch_tta_export --all --data-root data.
Evaluation conditions may be layered on top of a checkpoint via
load_checkpoint_from_hub(..., eval_condition="eqarch_tta_export") or
RuleInducer.from_pretrained("phuayj/gnri", eval_condition="eqarch_tta_export").
Checkpoints
Shipped checkpoints are the gnri training-condition seeds 42, 43, 44, 45, 46, 47, 48, 49. Verify
integrity against manifest_sha256.json:
import hashlib, json
from huggingface_hub import hf_hub_download
manifest = json.load(open(hf_hub_download("phuayj/gnri", "manifest_sha256.json")))
for entry in manifest["checkpoints"]:
digest = hashlib.sha256(open(hf_hub_download("phuayj/gnri", entry["path_in_repo"]), "rb").read()).hexdigest()
assert digest == entry["sha256"], entry["path_in_repo"]
- Downloads last month
- 26