freeqaz

decomp-synth-lifter-v17-mix-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 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 evaluation has been run on this adapter yet. It finished training on 2026-08-31 and is published as the freshly trained artifact; the project's multi-turn eval re-baseline is in flight, and sibling arms (pure-mint full/matched, r64, and 27B variants) train alongside it. Until numbers land, v14-evalformat remains the measured pick for the multi-turn feedback loop.
Training loss learned cleanly (0.77 → 0.24 over 1,602 optimizer steps, 2 epochs; min 0.02) — a supervision readout from the artifact's train_summary.json, not an evaluation claim.
Mix-vs-pure-mint is an unmeasured contrast. This arm exists to test whether restoring single-turn supply helps the trajectory generation; no verdict exists yet, and any future one is provisional under the project's multi-seed rule (single seed, 42).

Model facts

Table

base	`Qwen/Qwen3.5-9B` (apache-2.0)
method	LoRA r=32, α=64, dropout 0.05, all-linear, bf16
corpus	25,624 mixed rows (`v17mix.jsonl`, sha `b67811d1…`)
window	90,112 tokens max_seq (zero rows dropped)
schedule	2 epochs, 1,602 steps, eff. batch 32, linear LR, seed 42
trained	2026-08-31
format	Qwen3.5 chat template, thinking ON, native XML tool calls

Training data

The corpus mixes four layers. Counts are from the mix build's manifest (exact-dedup across layers, mint rows winning — 1,106 duplicate rows dropped):

Teacher-minted tool trajectories (3,032 rows) — the complete 2026-08-30/31 mint union, byte-identical to the corpus of the sibling v17-full arm (sha 411cfaed…): multi-turn sessions where a tool-armed teacher (GLM-5.3-flash) works decompilation targets from a private Halo: Combat Evolved Anniversary (Xbox 360) reconstruction effort through the project's compile-and-score harness — reading disassembly, proposing C++, compiling with the real toolchain, and reacting to objdiff feedback across turns.
bundle-v5 (8,305 rows) — single-turn repair, synthesis-from-stub, and edit-proposal rows from the project's verified-attempt bundle, the same source set as the v16-v5chat corpus: change histories of community decompilation projects — rb3-xenon, dc3-decomp, zeldaret/tww, zeldaret/tp, DarkRTA/rb3 — plus rows from the same private reconstruction effort, which is not publicly linkable. See each linked project for its own terms.

The single-turn layers were re-rendered through Qwen3.5-9B's own tokenizer template (thinking ON, with the same served-tools context the eval driver arms) and re-gated against the project's translation-unit-grain eval holdouts at mix build time. No license claim is made for the corpus; it is not published. The shipped PUBLISHED.json pins the corpus file's sha256 (b67811d1e00e…, 25,624 rows) 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-v17-mix-qwen3.5-9b-lora")

This adapter is trained for agentic tool use as well as single-turn proposal shapes: serve it behind an OpenAI-compatible endpoint with tool calling enabled and give it compile / diff / read tools. With vLLM:

markdown
vllm serve Qwen/Qwen3.5-9B --enable-lora \
  --lora-modules v17mix=freeqaz/decomp-synth-lifter-v17-mix-qwen3.5-9b-lora \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

The tool-call parser flag matters: the adapter emits tool calls in the Qwen3.5 template's native XML form, which qwen3_coder parses; without it (or with a JSON-style parser) the calls degrade silently into plain text.

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. Trajectory rows ran up to ~87k tokens of session context, so give it a long context window. In bf16 the 9B base fits a single 24 GB GPU.

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