pegasus-0.5-7b
uncc's own matching-decompilation model. Given one function's target assembly and its toolchain, it writes C meant to recompile to the identical bytes (relocations masked), not just to behave the same.
- base:
Qwen/Qwen2.5-Coder-7B-Instruct, LoRA SFT (r 64, alpha 128, 2 epochs, seq 8192), merged to bf16 - the LoRA adapter alone is in
adapter/ - rung r0: x86 (i386),
gcc 15.2 -O2, open-source C - site: uncc.sh · data: uncc.sh/dataset
Data
2,755 training pairs. Every pair is a function an uncc agent matched byte for byte on r0 targets (doom, zlib,
quake, duke3d, quake2): the server recompiled the agent's C with the pinned toolchain and got sha1: OK. Agent
claims are never trusted; only the server's re-check admits a pair. The targets' original source was never given to
the agents. Splits are by function (stable hash); the test set additionally holds out whole translation units, and
identical asm is deduplicated before splitting.
The full, growing pair set is public: uncc.sh/api/dataset/pairs.jsonl.
Eval
Every answer is compiled and diffed by the uncc engine with the same toolchain. Same prompt for every model; the
harness prepends the translation unit's #include lines to any answer without one, for every model alike.
- held-out: 234 r0 functions never seen in training
- unseen programs: lua 5.4 (200 fns) + brogue CE (200 fns), not in training at all
byte-exact on the first try (greedy):
| model | params | held-out | unseen |
|---|---|---|---|
| pegasus-0.5 | 7B | 23.1% | 6.3% |
| Qwen2.5-Coder-32B-Instruct | 32B | 9.0% | 4.3% |
| Qwen3-Coder-30B-A3B-Instruct | 30B MoE | 7.7% | 3.5% |
| DeepSeek-Coder-V2-Lite-Instruct | 16B MoE | 6.0% | 2.3% |
| Qwen2.5-Coder-7B-Instruct (base) | 7B | 5.1% | 2.5% |
byte-exact within 9 tries (greedy + 8 samples), every try checked by the compiler:
| model | held-out | unseen |
|---|---|---|
| pegasus-0.5 | 28.2% | 8.3% |
| Qwen2.5-Coder-32B-Instruct | 16.2% | 7.0% |
| Qwen3-Coder-30B-A3B-Instruct | 9.4% | 5.5% |
| DeepSeek-Coder-V2-Lite-Instruct | 7.7% | 4.5% |
| Qwen2.5-Coder-7B-Instruct (base) | 6.4% | 5.0% |
On programs it never saw the gain is small. That is what the next rungs and the next pegasus are for.
Use
The prompt format matters. prompt.py here is the exact template used in training (uncc-chat-v2):
from transformers import AutoTokenizer, AutoModelForCausalLM
from prompt import build_messages, extract_c
tok = AutoTokenizer.from_pretrained("uncc-sh/pegasus-0.5-7b")
model = AutoModelForCausalLM.from_pretrained("uncc-sh/pegasus-0.5-7b", torch_dtype="auto", device_map="auto")
pair = {"symbol": "...", "size": 0, "asm": "...", "toolchain": "gcc-i686-o2", "compiler": "gcc 15.2",
"flags": "-O2 -march=i686 ...", "arch": "i386"} # one row of pairs.jsonl, minus "c"
msgs = build_messages(pair, with_answer=False)
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(extract_c(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)))
A match only counts when the C compiles with the target's exact compiler and flags to the same bytes. Check it.
Ladder
r0 = gcc on open C (this model) → r1 = IDO · PsyQ · Borland (pegasus-1) → r2 = MSVC 5/6 → r3 = MWCC · ee-gcc · MSVC 7–9 → r4 = MSVC 14+, x64, LTO. Each rung's verified pairs train the next pegasus.
Licences of the r0 targets differ (GPL-2.0 for the id games and duke3d, zlib licence for zlib); check them before redistributing derived C.
- Downloads last month
- 37