cd /news/machine-learning/sly-cooper-and-the-thievius-raccoonu… · home › topics › machine-learning › article
[ARTICLE · art-173183] src=reversedwithai.com ↗ pub= topic=machine-learning verified=true sentiment=· neutral

Sly Cooper and the Thievius Raccoonus and Sly 2 (PS2): A local fine-tuned model that proposes matching C++ for PS2 functions, trained on the Sly Cooper games

A fine-tuned Gemma 4 E2B model with LoRA adapters, ps2-decomp-assist, proposes C++ for PS2 functions of at most 400 bytes written in MIPS R5900 assembly, and each proposal is rebuilt with the decompilation project's own compiler so only byte-identical candidates count as matches. Trained and measured on Sly Cooper and the Thievius Raccoonus and Sly 2 (NTSC-U), the model reached a 48.6 percent exact-match rate on functions up to 100 bytes at 8 samples across 150 held-out functions, with almost nothing matching above 200 bytes, and zero-shot Claude Haiku matched more functions than the local model in a 20-function comparison. The README states no disc data, assembly, dataset, compiler or SDK file is in the repository, and a guard test checks for this.

recommendation

Verify decompilation proposals by rebuilding them with the project's own compiler and counting only byte-identical matches, and benchmark against a frontier zero-shot model before trusting a small fine-tune.

more in playbook: Evals and testing for LLM apps
by read1 min views2 publishedNov 20, 2026
Sly Cooper and the Thievius Raccoonus and Sly 2 (PS2): A local fine-tuned model that proposes matching C++ for PS2 functions, trained on the Sly Cooper games
Image: Reversedwithai (auto-discovered)

What was done #

The model proposes C++ for PS2 functions of at most 400 bytes and each proposal is rebuilt with the decompilation project's own compiler. The README reports exact matches for part of 150 held-out functions.

ps2-decomp-assist is a fine-tuned Gemma 4 E2B model (LoRA adapters, MLX, Apple silicon) that proposes C++ for small PS2 functions written in MIPS R5900 assembly, at most 400 bytes each. Each proposal is checked with the decompilation project's own compiler and comparer, and only a candidate that rebuilds byte-identical and passes the project's rule checks is reported as a match. It was trained and measured on Sly Cooper and the Thievius Raccoonus and on Sly 2 (NTSC-U).

How it was done

The README says training pairs came from assembly generated from discs the authors own and from the authors' own decompiled source. It says no disc data, assembly, dataset, compiler or SDK file is in the repository, and a guard test checks for this. A Claude Code skill lets an agent call the command line tool, and Claude Sonnet 5.5 co-author trailers appear on the commits.

Results

On 150 held-out functions the README gives exact-match rates for the adapters against a few-shot baseline, and says functions up to 100 bytes reach 48.6 percent at 8 samples while almost nothing matches above 200 bytes. It also reports that zero-shot Claude Haiku matched more functions than the local model in a 20-function comparison.

Discussion #

Your library

Sources #

Sources reviewed · Added

An AI model drafted this entry from the sources above. An automated check matched its claims against them on 20 Nov 2026. The figures are the author's. How entries are checked. Something wrong? Tell me.

── more in #machine-learning 4 stories · sorted by recency
── more on @gemma 4 e2b 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/sly-cooper-and-the-t…] indexed:0 read:1min 2026-11-20 · —