Snowflake Proposes Excot: A Novel Ai Framework That Iteratively Optimizes Open-source Llms By Combining Cot Reasoning With Off-policy And On-policy Dpo, Relying Solely On Execution Accuracy As Feedback

Text-to-SQL translation, nan task of transforming earthy connection queries into system SQL statements, is basal for facilitating user-friendly database interactions. However, nan task involves important complexities, notably schema linking, handling compositional SQL syntax, and resolving ambiguities successful personification queries. While Large Language Models (LLMs) person shown robust capabilities crossed various domains, nan efficacy of system reasoning techniques specified arsenic Chain-of-Thought (CoT) wrong text-to-SQL contexts remains limited. Prior attempts employing zero-shot CoT aliases Direct Preference Optimization (DPO) without system reasoning yielded marginal improvements, indicating nan necessity for much rigorous methodologies.

Snowflake introduces ExCoT, a system model designed to optimize open-source LLMs done nan operation of CoT reasoning and iterative penchant optimization, specifically utilizing off-policy and on-policy DPO guided exclusively by execution accuracy feedback. ExCoT dispenses pinch outer reward models and quality annotations, relying alternatively connected internally generated reasoning steps and execution results. The method operates successful 2 main phases: initially, it generates campaigner CoT information validated done off-policy DPO, forming nan ground for supervised fine-tuning. Subsequently, nan exemplary iteratively generates and refines CoT information via on-policy DPO, incrementally improving accuracy done feedback derived from execution correctness.

ExCoT employs elaborate CoT reasoning, peculiarly adopting a divide-and-conquer strategy wherein analyzable queries are decomposed into simpler sub-queries. Each sub-query is analyzed and independently resolved earlier being integrated into a coherent last query. This system decomposition enables nan exemplary to negociate nan complexity and nested structures communal successful SQL operations much effectively. Execution-based verification serves arsenic nan halfway system for correctness evaluation, wherever generated queries are validated by comparing their execution outputs against ground-truth results. Incorrect and correct queries are systematically paired, providing definitive signals for preference-based learning. The iterative refinement successful nan on-policy DPO shape progressively enhances nan model’s reasoning accuracy.

Experimental information of ExCoT demonstrated important improvements successful execution accuracy. Specifically, pinch nan LLaMA-3.1 70B model, ExCoT elevated execution accuracy connected nan BIRD improvement group from 57.37% to 68.51%, and accrued Spider trial group capacity from 78.81% to 86.59%. Comparable capacity enhancements were recorded pinch nan Qwen-2.5-Coder 32B model. These results position ExCoT arsenic a starring attack successful single-model evaluations for these benchmarks, surpassing established methods specified arsenic XiYanSQL and proprietary models including OpenAI variants. Notably, nan improvements consistently maintained precocious query validity rates (exceeding 98%), confirming enhancements successful semantic correctness alongside syntactic precision.

In conclusion, ExCoT represents a methodical advancement successful system reasoning optimization for open-source LLMs applied to text-to-SQL tasks. By integrating system CoT reasoning pinch penchant optimization, guided solely by execution-based feedback, ExCoT efficaciously addresses limitations identified successful erstwhile methods. Its iterative refinement capacity ensures continuous betterment without dependence connected outer reward structures aliases manual annotations. Further investigation mightiness research extending this model to much intricate schema environments and further system reasoning tasks, frankincense broadening nan applicability and reliability of LLMs successful system query procreation contexts.

Check out the Paper, GitHub Page and Details. All in installments for this investigation goes to nan researchers of this project. Also, feel free to travel america on Twitter and don’t hide to subordinate our 85k+ ML SubReddit.

🔥 [Register Now] miniCON Virtual Conference connected OPEN SOURCE AI: FREE REGISTRATION + Certificate of Attendance + 3 Hour Short Event (April 12, 9 am- 12 p.m. PST) + Hands connected Workshop [Sponsored]