Turing agro LLM13

Client

Turing product line - offered to mill groups and to an agriscience company as a custom agent over internal documents

Industry

Agriculture · Sugar & ethanol · Seed and crop protection

Problem

General-purpose models answer agricultural questions fluently and often wrongly; nobody's internal reports are in their training data

Solution

A domain model and knowledge base: a crawler over the Brazilian agro web and document corpus, a fine-tuned open model with DPO refinement measured by semantic entropy, vLLM serving on a single GPU, long-document summarization, an orchestrated data pipeline, and a fine-tuning path that turns a client's internal documents into a private agent

Stack

Custom crawler (GitHub Actions checks, graceful stop, multi-format downloads) · S3 bronze/silver/gold layers · Prefect orchestration + experiment tracking · fine-tuning on ~450k examples · DPO · semantic entropy evaluation · Phi-4 base for summarization · vLLM on NVIDIA L40S · Kubernetes · Argo CD deployment

Live at

app.turing.com.br (demo by registration)

Ask a frontier model about Brazilian tilapia exports, regional sugarcane productivity, or the latest crop-insurance statistics by municipality and it will answer. The answer is a plausible paragraph with a good chance of a wrong number. The gap is not intelligence; it is that the documents holding those numbers are scattered across CONAB bulletins, Embrapa reports, sector PDFs and trade press, and were never assembled into anything a model could be grounded on.

1. 01

   The corpus

   A crawler walks a curated list of base domains (base_urls.json), downloads PDFs and, after a fix, every other useful format (csv, doc/docx, epub, ods, ppt/pptx, rtf, tsv, txt, xls/xlsx). Broken domains fail an automatic check integrated with GitHub Actions and runnable locally with make check. Visited pages are saved each round, and the crawler stops gracefully so no file is left corrupted. Cleaning rules: drop files that are mostly XML/HTML tags rather than stripping tags, drop files under 256 tokens, and cap downloads (spreadsheets of 1 GB+ were found in the wild). The corpus covers productivity by region, financial assessments and projections, government plans, crop-loss statistics by locality, and commodity prices, among other topics.

2. 02

   The pipeline

   Runs are orchestrated with Prefect and tracked as experiments. Three jobs: crawl to a bronze S3 bucket per run (crawler/runs/run_YYYYMMDD_HHMMSS), clean to silver, and rebuild the knowledge-base index to gold, merging the crawl with the books-and-papers and reports buckets. Crawler metadata is stored at indexing time to refine search and, later, cite sources.

   

3. 03

   Fewer wrong numbers than GPT-4

   Domain questions with known answers were put to the fine-tuned model and to GPT-4. The Turing model returned the right figures more often, with less confident filler.

   

   

4. 04

   Fine-tuning and how it was measured

   The latest fine-tuning run used ~450,000 training examples: articles, books and web pages. Semantic entropy of model answers was the metric: lower means more consistent answers to rephrased questions. Version - Mean semantic entropy; Base model - 1.86; DPO I (demo version) - 1.58; Epoch 1 - 1.38; Epoch 2 - 1.41; Epoch 3 - 1.38; DPO II (instruction-tuned from epoch 3) - 1.41 Two findings were recorded rather than hidden: from the second epoch the model began reproducing training-data noise at the end of long texts (the same behavior seen in the 200k-example run), and DPO II underperformed because its preference data was stale. The fix is a refreshed, larger preference dataset for the next refinement.

   

5. 05

   Serving

   Inference runs on vLLM. On a single NVIDIA L40S it supports more than 10 simultaneous users with stable latency; request scheduling absorbs bursts as a controlled latency increase rather than out-of-memory failures. Load tests were added, and multi-instance tests on Kubernetes were next. The demo ran on a g6e.2xlarge because of the memory the indexed corpus requires.

6. 06

   Long documents

   The summarization base (Phi-4) has a 16K-token window. Long files are split into n chunks, each summarized, and a final summary is generated from the concatenated chunk summaries. Information loss grows with document size and is declared; a 150-page report summarizes in a little over two minutes.

   

   

7. 07

   A private agent on a client's documents

   The same model is the base for a client-specific agent. The agro-tuned Turing LLM is fine-tuned again on the client's internal documents, producing an agent that knows both the sector and the company. This is the configuration proposed to an agriscience R&D team for its internal report archive.

   

8. 08

   Deployment

   Deployment pipelines are fully automated with Argo CD; a merged change ships itself.