MagicNoThief/cs2-overwatch

CS2 Overwatch review: models

The two models behind Overwatch review, an offline reviewer for Counter-Strike 2 demos that runs on your own PC. The app downloads these files itself, pinned to one commit and checked against their SHA-256; you do not need to fetch them by hand.

file	what it is
detector/scorer.onnx	a 1D CNN that scores each kill's aim trajectory (136 KB)
detector/scorer.json	its architecture and the frozen clean-player reference a score is read against
judge/judge-v3.Q4_K_M.gguf	Qwen3.5-4B fine-tuned (QLoRA) to write a verdict from the evidence, Q4_K_M

What they are for

To help a person review a demo: which players and which kills deserve a look, and why, in terms that can be checked in the demo. Not to ban anyone automatically. A high score says a player's kills look unlike clean players' kills in the training data; it is evidence to examine, not proof.

How they were made

• Detector: trained on CS2CD (795 matches, 317 with a VAC-banned player); per-player ROC-AUC 0.93 in match-grouped cross-validation. Line of sight comes from ray casts against each map's collision mesh, not the game's spotting flag, which is biased against snipers.
• Judge (v3): fine-tuned on generated verdicts whose targets depend only on evidence shown in the text (never on the ban label), with clean players' 95th and 99th percentiles printed beside every measurement. On held-out cases it matches its targets 87% of the time, accuses 1.9% of clean players, and cited no number absent from the evidence in 206 answers.

Credits and licences

• These models are released under CC BY-NC 4.0: free to use, share and adapt for non-commercial purposes, with credit. The app's code is AGPL-3.0-or-later.
• The judge is a fine-tune of Qwen3.5-4B by the Qwen team, licensed under Apache-2.0; its licence is included as judge/LICENSE-Qwen3.5-Apache-2.0.txt.
• Both models were trained on the CS2CD dataset (CC-BY-4.0) by Mille Mei Zhen Loo and Gert Lužkov.