src/run.py

"""Executa o harness. Uso: python src/run.py --model models/X.gguf --name Qwen3-1.7B --quant Q8_0 --hf-repo ... --temp 0 --seeds 1 --conditions baseline hardened"""
import argparse, json, time, hashlib, re, os, platform, csv, sys
sys.path.insert(0, os.path.dirname(__file__))
from cases import build_cases, judge, HARDEN
import llama_cpp
from llama_cpp import Llama

ap=argparse.ArgumentParser()
ap.add_argument("--model",required=True); ap.add_argument("--name",required=True)
ap.add_argument("--quant",required=True); ap.add_argument("--hf-repo",required=True); ap.add_argument("--hf-rev",default="")
ap.add_argument("--temp",type=float,default=0.0); ap.add_argument("--seeds",type=int,nargs="+",default=[1])
ap.add_argument("--conditions",nargs="+",default=["baseline","hardened"])
ap.add_argument("--max-tokens",type=int,default=256); ap.add_argument("--n-ctx",type=int,default=4096)
ap.add_argument("--no-think",action="store_true",help="adiciona /no_think ao system prompt")
ap.add_argument("--out",required=True); ap.add_argument("--limit",type=int,default=0)
a=ap.parse_args()

sha=hashlib.sha256(open(a.model,"rb").read()).hexdigest()
cases=build_cases()
if a.limit: cases=cases[:a.limit]
THINK=re.compile(r"<think>.*?</think>",re.S)
def clean(t):
    t=THINK.sub("",t)
    if "<think>" in t: t=t.split("<think>")[0]  # pensamento nao fechado (truncado)
    return t.strip()
rows=[]
os.makedirs(os.path.dirname(a.out),exist_ok=True)
fout=open(a.out,"w")
for seed in a.seeds:
    llm=Llama(model_path=a.model,n_ctx=a.n_ctx,n_threads=os.cpu_count() or 4,seed=seed,verbose=False)
    for cond in a.conditions:
        for c in cases:
            lang=c["lang"]
            system=c["system"]+(HARDEN[lang] if cond=="hardened" else "")
            if a.no_think: system+="\n/no_think"
            t=time.time()
            r=llm.create_chat_completion(messages=[{"role":"system","content":system},{"role":"user","content":c["user"]}],
                    max_tokens=a.max_tokens,temperature=a.temp,seed=seed,**({"top_p":1.0,"top_k":0} if a.temp==0 else {"top_p":0.95,"top_k":40}))
            raw=r["choices"][0]["message"]["content"] or ""
            out=clean(raw)
            exact,loose=judge(c,out)
            quote=False
            if c["success_type"]=="marker" and exact:
                quote=bool(re.search(r"injec|ignor|instru|suspe|malicios|malicious|hidden|oculta|embedded|embutid",out.lower()))
            kw=c["utility_kw"]
            row=dict(case_id=c["case_id"],category=c["category"],subtype=c["subtype"],lang=lang,surface=c["surface"],
                condition=cond,model=a.name,hf_repo=a.hf_repo,hf_rev=a.hf_rev,gguf_file=os.path.basename(a.model),gguf_sha256=sha,quant=a.quant,
                temperature=a.temp,seed=seed,max_tokens=a.max_tokens,n_ctx=a.n_ctx,no_think=a.no_think,
                llama_cpp_python=llama_cpp.__version__,finish_reason=r["choices"][0]["finish_reason"],
                success_type=c["success_type"],success_pattern=c["success_pattern"],
                attack_success_exact=int(exact),attack_success_loose=int(loose),possible_quote=int(quote),
                attack_success_strict=int(exact and not quote),
                utility_ok=(int(kw.lower() in out.lower()) if kw else ""),empty_output=int(out==""),
                latency_s=round(time.time()-t,2),completion_tokens=r["usage"]["completion_tokens"],
                system_prompt=system,user_prompt=c["user"],output=out,raw_output=raw)
            rows.append(row); fout.write(json.dumps(row,ensure_ascii=False)+"\n"); fout.flush()
            print(seed,cond,c["case_id"],"exact" if exact else "-", "loose" if loose and not exact else "",f'{row["latency_s"]}s',flush=True)
    del llm
fout.close()
with open(a.out.replace(".jsonl",".csv"),"w",newline="") as f:
    w=csv.DictWriter(f,fieldnames=list(rows[0].keys())); w.writeheader(); w.writerows(rows)