A 4B model that ties GPT-5.5 on financial question answering and shows its receipts: it cites figures from 10-K filings, a calculator computes, and a verifier checks every number. SFT and GRPO with verifiable rewards, vLLM serving, adversarial testing, 299 tests.
reinforcement-learning slurm pytorch fintech question-answering lora fine-tuning explainable-ai mlops sft huggingface llm rlhf vllm llm-evaluation hallucination-detection grpo rlvr finqa fianancial-nlp
-
Updated
Oct 11, 2026 - Python