Spaces:
Running
Running
| """ | |
| Default configuration for FactEval models and parameters. | |
| """ | |
| # ββ Model IDs (Hugging Face Hub) βββββββββββββββββββββββββββββββββββββββββββββ | |
| # Claim extraction β chosen in Week 0: 1.5B was 3.5x faster with cleaner output | |
| CLAIM_MODEL = "Qwen/Qwen2.5-1.5B-Instruct" | |
| # Sentence embeddings for evidence retrieval | |
| EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2" | |
| # NLI verification (used in Week 2) | |
| NLI_MODEL = "MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli" | |
| # ββ Retrieval defaults βββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| DEFAULT_TOP_K = 3 | |
| MIN_EVIDENCE_SCORE = 0.3 # Below this, evidence is too weak to use | |
| # ββ Claim extraction defaults ββββββββββββββββββββββββββββββββββββββββββββββββ | |
| MAX_NEW_TOKENS = 200 | |
| MAX_CLAIMS = 10 | |
| CLAIM_SYSTEM_PROMPT = ( | |
| "You are a claim extraction engine. Given a text, break it into atomic, " | |
| "independently verifiable claims. Each claim states exactly ONE fact. " | |
| "Return ONLY a numbered list. No explanations, no commentary." | |
| ) | |
| CLAIM_USER_PROMPT = "Break this into atomic claims:\n\n{text}" | |