Luimas/claim-extractor-qwen3b

Claim Extractor (Qwen2.5-3B, QLoRA, distilled from 7B)

Fine-tuned to read English text and emit strict JSON: a summary, publication date (if present), keywords, typed/categorized claims with verbatim evidence spans, and contradictions between claims. Built for rumor / fact-checking pipelines.

Output schema

{"summary": "...", "publication_date": "ISO date or null", "keywords": ["..."],
 "claims": [{"id": 0, "claim": "...", "claim_type": "fact|statistic|opinion|prediction|speculation|rhetoric|other",
              "category": "...", "importance": "high|medium|low",
              "stance": "asserted|denied|hedged|attributed", "evidence_span": "verbatim", "confidence": 0.0}],
 "contradictions": [{"claim_a": 0, "claim_b": 1, "relation": "contradiction|tension", "explanation": "..."}]}

Files

• merged_16bit/ — full fp16 model (HF format)
• Qwen2.5-3B-Instruct.Q4_K_M.gguf — Q4_K_M GGUF for llama.cpp (runs on a 4 GB GPU)
• lora_adapter/ — LoRA adapter only
• claim.gbnf + prompt.txt — grammar + instruction for guaranteed-valid JSON via llama-cpp-python

Quick use (llama.cpp, grammar-constrained → always valid JSON)

pip install llama-cpp-python
# load Qwen2.5-3B-Instruct.Q4_K_M.gguf with claim.gbnf as the grammar and prompt.txt prepended to your text

Training

Distilled from unsloth/Qwen2.5-7B-Instruct-bnb-4bit. Data: ag_news (teacher-labeled) + SNLI (contradiction pairs) + LIAR (misinformation) + teacher-synthesized adversarial passages. QLoRA r=32, 3 epochs.

Benchmarks

{
  "contradiction_recall": 0.75,
  "evidence_verbatim_rate": 1.0,
  "date_extraction_rate": 1.0
}
held_out_validity = 1.0, unseen_validity = 1.0