---
license: mit
library_name: sample-factory
tags:
- sample-factory
- reinforcement-learning
- doom_health_gathering_supreme
model-index:
- name: APPO-doom_health_gathering_supreme
  results:
  - task:
      type: reinforcement-learning
      name: reinforcement-learning
    dataset:
      name: doom_health_gathering_supreme
      type: doom_health_gathering_supreme
    metrics:
    - type: mean_reward
      value: 11.787799999999942 +/- 5.578349501420757
      name: mean_reward
---
# APPO Doom Health Gathering Supreme

Trained from random initialization by Kay Zheng (bestdive), with AI coding and execution assistance, for the Hugging Face Deep RL course Unit 8 Part II. No pretrained policy was copied and no evaluation metric was fabricated. Training used a free Colab T4 GPU, seed 42 and 4,005,888 environment frames.

## Independent evaluation

100 complete episodes in a separately seeded environment (100001), deterministic actions, latest final checkpoint. Mean reward 11.787799999999942, population standard deviation 5.578349501420757. The course score is 6.209450498579185 (required minimum 5).

These are the **true-objective episode returns** collected in Sample Factory's official enjoyment/evaluation function and used by its model-card generator. We do not substitute the higher shaped training reward. Every return is in evaluation.json. The exact checkpoint loaded is independently recorded in doom-evaluation.log.

## Reproduction

Python 3.10.12, sample-factory==2.1.1, vizdoom==1.2.4, torch==2.2.2, numpy<2, setuptools<81. Registration and training code comes from the official Hugging Face Unit 8 Part II notebook, with explicit experiment arguments. train_doom.py and evaluate_doom.py contain the executed code.

For inference, place this repository under /content/doom-results/Kay-Doom, copy train_doom.py and evaluate_doom.py into /content, then run evaluate_doom.py. The checkpoint must be in checkpoint_p0/. Training from scratch runs train_doom.py. Change the explicit paths in both scripts for another workspace.

This model is limited to this simulated environment and is not evidence of general AI competence.