freeqaz

decomp-synth-lifter-v16-v5chat-qwen3.5-9b-lora

README

License: apache-2.0

The project

This adapter comes out of a campaign teaching fleets of AI coding agents to reconstruct shipped game binaries byte-for-byte — matching decompilation, the console-preservation community's gold standard: rewrite C++ until the original vintage compiler, at the original flags, reproduces the retail machine code exactly. In that campaign, agent swarms byte-matched over 44,000 functions of a stripped retail Xbox 360 binary in about ten weeks — with every proposal judged by the real compiler and objdiff, never by a model.

decomp-synth is the search-and-ML side of that system: an engine that searches behavior-preserving C++ rewrites against the real build, and mines its own verified attempts into training data. This adapter is the broadest cut of that flywheel so far — and part of the project's practice of publishing research artifacts with their evidence stated plainly, including the null results. The design rule throughout: models propose, deterministic oracles dispose — byte-exact recompilation is the only terminal judge.

Background reading, in order:

How agent swarms decompile games byte-for-byte — the campaign, and what byte-exact matching actually demands
LLM-assisted decompilation: tools, not prompts — why the model is never allowed to judge its own work
Forking the decompilation toolchain — the compiler, splitter, and diff stack underneath
Infrastructure for running agent swarms — what running the fleet takes

Honest status notes

Published as a research artifact with its evidence stated plainly:

No measured win over v14. In a paired 4-model run on the multi-turn feedback instrument (118 start states, K=2 replicas), v16 does not separate from v14 at any rank: witness deltas +2/+1/+2, p = 1.0 / 1.0 / .77.
The rank ladder is flat. r32/r64/r128 trained on the byte-identical corpus measure within noise of each other (r64−r32 −1, p=.77; r128−r32 exactly 0, p=1.0), replicating a flat NLL ladder — adapter capacity is not the 9B ceiling, which is why this release is the r32.
Single training seed (seed=42, all rungs) — even the flat verdict is provisional on magnitude under the project's multi-seed rule.
The corpus is 100% single-turn while the target instrument is 4-turn — a known mismatch; the multi-turn supply line is separate, in-flight work.
The instrument itself was live in the same cells: real compiler feedback beat derangement-scrambled feedback in all six arm×replica cells (strongest: +8 discordant vs 0, p=.0078).

What it does bring over v14: ~2.2× the training rows, the proposal-model and synthesis task shapes, and broader substrate coverage (below).

Model facts

Table

base	`Qwen/Qwen3.5-9B` (apache-2.0)
method	LoRA r=32, α=64, dropout 0.05, bf16
corpus	8,735 chat-rendered rows (`v16_v5chat.jsonl`, sha `dc2f1a4e…`)
trained	2026-08-28
format	chat (Qwen chat template)

Training sources

Rows are derived from the change histories of these community decompilation projects (measured per-row attribution of the corpus):

rb3-xenon and dc3-decomp — Xbox 360 MSVC (≈7,600 rows)
zeldaret/tww and zeldaret/tp — GameCube mwcc (≈570 rows)
DarkRTA/rb3 — Wii mwcc (≈510 rows)
plus ≈1,260 rows from a private Halo: Combat Evolved Anniversary (Xbox 360) reconstruction effort, not publicly linkable

See each linked project for its own terms. The corpora themselves are not published; the shipped PUBLISHED.json pins the corpus file's sha256 as provenance.

How to use it

python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")
model = PeftModel.from_pretrained(base, "freeqaz/decomp-synth-lifter-v16-v5chat-qwen3.5-9b-lora")

Chat template. Prompts follow the project's repair / synthesis / proposal formats; like every adapter in this family, it is a proposal engine for a compile-and-score loop — sample multiple candidates, compile each with the target project's real toolchain, and let objdiff judge. Byte-exact is the only acceptance criterion. If your use case is the 4-turn feedback loop, prefer v14-evalformat (no measured v16 win there); reach for v16 when you want the synthesis-from-stub and edit-proposal task shapes or the broader substrate coverage. At 9B it fits a single 24 GB GPU in bf16; vLLM with --enable-lora serves it well.

The adapter family

Table with columns: adapter, base, format, role
adapter	base	format	role
v10	Qwen3.6-27B	RAW	single-turn matcher, reference 27B
gtp0c	Qwen3.8-27B	RAW	v10 recipe verbatim on the newer base (matched pair)
v13-chain r32	Qwen3.5-9B

Integrity

PUBLISHED.json (shipped) carries per-file sha256 written by the training job at publish time. Two internal metadata files it pins (train_summary.json, artifact-manifest.json) are retained in the project's artifact store but not shipped here — they carry training-box paths. adapter_config.json was rewritten post-training to the HF base id; original bytes at provenance/adapter_config.orig.json.

License

Adapter weights: Apache-2.0. A model trained to reproduce shipped game code can reproduce shipped game code — that is its purpose; decompilation projects operate in a long-standing community norm of clean-room reconstruction for interoperability and preservation.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

freeqaz

Model Tree

Base

Qwen/Qwen3.5-9B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality