仓衡大模型

Canbal v1.0 · Agricultural Intelligence System

A vertical-domain large language model system designed for smart agriculture. Built upon Qwen3-8B with extensive agricultural instruction tuning, the system integrates multi-source perception with intelligent decision-making. Pioneering the introduction of metacognitive mechanisms, the model achieves self-reflection and strategy optimization capabilities. It incorporates real-time information retrieval, meteorological data fusion, multi-agent collaborative reasoning, domain knowledge graph, and RAG-enhanced retrieval modules, providing comprehensive intelligent support for agricultural production, research, and management.

Highlights

Real-time retrieval · Meteorological fusion · Multi-agent reasoning · Metacognitive self-reflection · Domain knowledge graph · RAG-enhanced retrieval

Qwen3-8B + Large-scale SFT · Smart Agriculture Domain

Technical Report

This report currently contains the system architecture diagram. Detailed evaluation results are not yet available and will be updated here once released.

Architecture Overview

• S1 Data FoundationSFT dataset of 15,002 samples across 36 sub-categories; agronomy corpus (cropping · disease · soil); 646 de-duplicated policy documents; 2,487 curated and sourced encyclopedia entries; crop–disease–pesticide knowledge-graph triples; Chroma vector store with bge + BM25 index.
• S2 Domain ModelQwen3-8B + LoRA (r16 / α32); 4-bit QLoRA loading, single-GPU trainable; 15,002 supervised samples with assistant-only loss; KV int8 quantization saving ~50% KV memory; low-temperature decoding (0.1, top_p 0.9); repeat stop and de-duplication.
• S3 Orchestration & ToolsAgent ReAct with at most 5 steps (tools before answer); intent routing (multi-hop → agent); hybrid retrieval (3-way recall · rerank); knowledge-graph multi-hop (networkx · 2–3 hops); deterministic routing with exact dosage and units; metacognition check (confidence + logit gap).
• S4 World ModelGDD phenology (daily thermal-time accumulation); FAO-56 water balance (ET0 + crop coefficient); disease-risk index (weather × growth stage); spray window (suitable weather slot); locked default parameter layer (wm_params.json, read-only); calibration override (wm_calibrated.json).
• S5 Output & FeedbackSource citation [N] with source-level grading; confidence grading (low → clarification); pre-harvest interval safety notice shown once; expandable chain-of-thought (4-step reasoning); structured report generation; side-by-side crop table comparison.

External Interfaces

Base model Qwen3-8B-Instruct · QWeather API · Tavily web search · institution CSV data

Calibration Loop Institution field data (yield / phenology / disease / weather) → override layer hot-applied → the world model improves with use.