Benign huggingface Download

VertexAGI/experiment-rh1-reward-hacking main

Module built on the numeric Python stackImports the NumPy array library

SHA-25649c048a05104321e0db04decddc550084b50ed4e55f7b0c847983535658e0775

URLhttps://huggingface.co/VertexAGI/experiment-rh1-reward-hacking

Evidence

English function-word token "the" lines 1–5

1#!/usr/bin/env python3
2"""Experiment RH1: The Spilling Cleaner -- render one full episode of the
3trained policy to show the reward hack directly: the action sequence, and
4a heatmap of how many timesteps the agent spent in each grid cell (a
5genuinely task-completing  …

Imports the NumPy array library lines 6–22

⋯4 lines
10
11import mlx.core as mx
12import numpy as np
13
14from env import SpillingCleanerEnv, ACTIONS
⋯8 lines

NumPy array operation (zeros) lines 23–29

24… .obs_dim, len(ACTIONS))
  policy.load_weights(os.path.join(HERE, "policy.safetensors"))
  mx.eval(policy.parameters())
26
  visit_counts = np.zeros((env.grid_size, env.grid_size), dtype=int)
  obs = env.reset()
  print(f"Initial dirty cells: {int(env.dirty.sum())} / {env …

JSON command or task field key lines 30–37

30:24… ion: {tuple(env.pos)}\n")
31    print(f"{'step':>4}  {'pos':>6}  {'action':>7}  {'reward':>6}  {'note'}")
32
33    done = False
⋯4 lines

English function-word token "this" lines 45–47

10…    # NOTE: was_dirty_here reflects state BEFORE this step's action;
          # the self-spilled/genuine distinction is tracked precisely
       …