2026-11-16 · America/Los_Angeles · 社区动态 · #10
GOOD LUCK
Başlık: Small local models + a code-side "gate": 54% → 92% accuracy and 0 silent wrong answers on table tasks (informal test, raw data public) Metin: We ran an informal A/B test on three local Qwen2.5 models (1.5B, 3B, 7B GGUF, llama.cpp `llama-server`, CPU only) on simple table questions (sum, average, min/max, filters). - Arm A: plain LLM writes Python, we run it. - Arm B: the model writes a tiny command language (DSL). Code parses it, a gate checks the command against a catalog, and only then a deterministic runner computes the result. Out-of-catalog requests are rejected with a reason.…
热度 40.8 / 100;排名与评分保留该期记录。