初始化项目，由ModelHub XC社区提供模型

Model: Luimas/claim-extractor-detective-qwen3b
Source: Original Platform

scripts/evaluate.py

@@ -0,0 +1,26 @@
+#!/usr/bin/env python
+# Quick self-contained eval: runs the GGUF (grammar-constrained) on built-in test passages and
+# reports JSON-validity, sarcasm handling, and contradiction detection.  pip install llama-cpp-python
+import glob, json, os
+from llama_cpp import Llama, LlamaGrammar
+HERE = os.path.dirname(os.path.abspath(__file__)); ROOT = os.path.dirname(HERE)
+GGUF = (glob.glob(os.path.join(ROOT, "*.gguf")) + glob.glob(os.path.join(HERE, "*.gguf")))[0]
+PROMPT = open(os.path.join(ROOT, "prompt.txt"), encoding="utf-8").read()
+GRAMMAR = LlamaGrammar.from_string(open(os.path.join(ROOT, "claim.gbnf"), encoding="utf-8").read())
+CASES = [
+    ("The mayor said crime fell last year. Hours later the police chief said it actually rose.", "contradiction"),
+    ("What an incredible update, my phone now dies by noon. The engineers really outdid themselves.", "sarcasm"),
+    ("On 2024-03-12 the central bank cut rates to 4.5%, citing easing inflation.", "fact"),
+]
+llm = Llama(model_path=GGUF, n_ctx=4096, n_gpu_layers=-1, verbose=False)
+ok = con = sar = nclaims = 0
+for text, tag in CASES:
+    o = llm.create_chat_completion(messages=[{"role": "user", "content": PROMPT + text}],
+                                   grammar=GRAMMAR, temperature=0.0, max_tokens=768)
+    try:
+        d = json.loads(o["choices"][0]["message"]["content"]); ok += 1; nclaims += len(d.get("claims", []))
+        if tag == "contradiction" and d.get("contradictions"): con += 1
+        if tag == "sarcasm" and any(c.get("stance") == "ironic" for c in d.get("claims", [])): sar += 1
+    except Exception:
+        pass
+print(f"valid JSON: {ok}/{len(CASES)} | avg claims: {nclaims/max(1,ok):.1f} | contradiction: {con}/1 | sarcasm: {sar}/1")

scripts/inference.py

@@ -0,0 +1,24 @@
+#!/usr/bin/env python
+# Local grammar-constrained inference (always-valid JSON).  pip install llama-cpp-python
+import argparse, glob, json, os, sys
+from llama_cpp import Llama, LlamaGrammar
+HERE = os.path.dirname(os.path.abspath(__file__)); ROOT = os.path.dirname(HERE)
+def _find(name):
+    for d in (ROOT, HERE):
+        hits = glob.glob(os.path.join(d, name))
+        if hits: return hits[0]
+    raise FileNotFoundError(name)
+GGUF = _find("*.gguf")
+PROMPT = open(_find("prompt.txt"), encoding="utf-8").read()
+GRAMMAR = LlamaGrammar.from_string(open(_find("claim.gbnf"), encoding="utf-8").read())
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--text"); ap.add_argument("--file"); ap.add_argument("--n_gpu_layers", type=int, default=-1)
+    a = ap.parse_args()
+    text = open(a.file, encoding="utf-8").read() if a.file else (a.text or sys.stdin.read())
+    llm = Llama(model_path=GGUF, n_ctx=4096, n_gpu_layers=a.n_gpu_layers, verbose=False)
+    out = llm.create_chat_completion(messages=[{"role": "user", "content": PROMPT + text[:4000]}],
+                                     grammar=GRAMMAR, temperature=0.0, max_tokens=768)
+    print(json.dumps(json.loads(out["choices"][0]["message"]["content"]), indent=2, ensure_ascii=False))
+if __name__ == "__main__":
+    main()

scripts/inference_hf.py

@@ -0,0 +1,23 @@
+#!/usr/bin/env python
+# Inference via the merged fp16 model (Transformers).  pip install transformers accelerate torch
+import argparse, json, os, sys
+from transformers import AutoModelForCausalLM, AutoTokenizer
+HERE = os.path.dirname(os.path.abspath(__file__)); ROOT = os.path.dirname(HERE)
+PROMPT = open(os.path.join(ROOT, "prompt.txt"), encoding="utf-8").read()
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--text"); ap.add_argument("--file"); ap.add_argument("--model", default=ROOT)
+    a = ap.parse_args()
+    text = open(a.file, encoding="utf-8").read() if a.file else (a.text or sys.stdin.read())
+    tok = AutoTokenizer.from_pretrained(a.model)
+    model = AutoModelForCausalLM.from_pretrained(a.model, torch_dtype="auto", device_map="auto")
+    ids = tok.apply_chat_template([{"role": "user", "content": PROMPT + text[:4000]}],
+                                  return_tensors="pt", add_generation_prompt=True).to(model.device)
+    out = model.generate(ids, max_new_tokens=768, do_sample=False, pad_token_id=tok.eos_token_id)
+    raw = tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True)
+    try:
+        print(json.dumps(json.loads(raw[raw.find("{"):raw.rfind("}") + 1]), indent=2, ensure_ascii=False))
+    except Exception:
+        print(raw)
+if __name__ == "__main__":
+    main()