Safety routing language model

Frozen language-model checkpoint used to generate the router-probability safety features analyzed in report.md in the latentreasoning project. This is the feature-producing language model; the downstream safety classifiers and datasets are not included.

  • Architecture: mol_dynamic_window, 33 experts, 22 routing depths.
  • Training run: qz0v9938.
  • Checkpoint step: 9389.
  • File: latest_ce6a3160ded0.pt (4,831,570,306 bytes).
  • SHA-256: 284a60f0ce565ec2d5d7bf09bc3e8f64d6f64588401af7325724b06e6465e9dd.

The checkpoint is preserved byte-for-byte in its original native PyTorch format. It requires the latentreasoning project code and environment; it is not a Transformers from_pretrained export. The accompanying provenance.json records the study configuration, software versions, and provenance. Paths in that file refer to the original research workspace.

Loading

Run from the latentreasoning project environment:

import json
import torch
from huggingface_hub import hf_hub_download
from evaluation.model_loading import build_evaluation_model

repo_id = "smcleish/latentreasoning-safety-routing"
checkpoint_path = hf_hub_download(repo_id, "latest_ce6a3160ded0.pt")
provenance_path = hf_hub_download(repo_id, "provenance.json")
with open(provenance_path) as stream:
    provenance = json.load(stream)
checkpoint = torch.load(checkpoint_path, map_location="cpu", mmap=True, weights_only=False)
model, _ = build_evaluation_model(checkpoint["model_state_dict"], provenance["config"])
model.eval()

This follows the study's loading path. Router feature extraction additionally uses the project's causal window and document-boundary handling. The report defines each feature as a router probability for one of 22 depths and 33 experts; downstream classifier metrics are not language-model safety guarantees.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support