Reasoning SLM for agriculture: research preview

On 14 May 2025, we released two experimental models for Climate-Resilient and Sustainable Agriculture in India. ThinkingDhenu1-CRSA-India-preview is a 4-billion-parameter model based on Qwen3-4B. The micro preview is based on Qwen3-0.6B.

The current model cards are the source of record for the base models, training method, licence, and limitations. This post explains what we were trying to learn from the preview and corrects some details from the original release note.

What reasoning means in this preview

The training samples use a chain-of-thought plus answer format. Each dialogue separates a private <think> section from the final answer. That structure gives the model room to work through conditions before it responds, while reducing the chance that a raw reasoning trace appears in the user-facing answer.

That distinction matters in agriculture. A useful answer about pigeon pea may depend on rainfall expectations, soil moisture, labour availability, crop rotation, and local prices. The model has to account for those conditions without presenting its private trace as an agronomist’s verbatim reasoning or as proof that the answer is correct.

The final answer still needs grounding and review. The model cards explicitly rule out real-time critical decisions without human validation.

What we trained on

The Thinking-climate-100k dataset contains about 101,000 synthetic, aligned, multi-turn dialogues on climate-smart agriculture. The examples were generated from the topics, language patterns, and decision contexts found in real agricultural conversations. They are not verbatim farmer queries, and the dataset should not be described as a collection of hand-written traces from individual farmers or outside agronomists.

KissanAI’s internal agronomy team curated the source material, reviewed difficult samples, corrected regional and crop-specific errors, and used those corrections to improve the generation and filtering rules. Agronomist time remained the expensive part. A difficult trace can take 20 to 30 minutes to review and correct, even though the wider dataset is generated synthetically.

This is what we mean by “synthetic and aligned from real conversations”: real conversations inform the subject matter and failure cases, synthetic generation creates the training examples, and internal agronomy review shapes the alignment process.

The base models and training method

The 4B preview uses Qwen/Qwen3-4B. The micro preview uses Qwen/Qwen3-0.6B. Both are Qwen3 decoder-only causal language models with a 32K context window. We used supervised fine-tuning through LLaMA-Factory. The released model weights inherit the Apache 2.0 licence from the base models.

The micro model is intended for experiments where memory and inference cost matter. We have not published a controlled device-latency benchmark, so the release note does not make a sub-second phone claim.

What the internal evaluation tells us

Our private review set indicated that the 4B preview performed better than the larger Dhenu2 8B model on the narrow CRSA tasks used for this experiment. That result is directional. The public model card does not currently include the private evaluation numbers, and readers should not treat the comparison as an independently reproduced benchmark.

Public evaluation is the next useful step. It should cover answer quality, grounding, regional validity, safety, and how often the model asks for missing context rather than guessing.

Scope and limitations

The preview is narrow by design. It focuses on climate-resilient and sustainable agriculture in India, including climate-smart cropping, pest IPM, soil and nutrient management, and natural-farming practices.

The model cards also name the current limitations:

• The models primarily handle English and technical Indian agricultural vocabulary.
• The dataset may carry a bias toward Andhra Pradesh Community Managed Natural Farming practices.
• Climate data cited in the training material is static as of 2024 and must be checked against current IMD advisories.
• The models can hallucinate and require qualified human validation for high-stakes advice.

What is available

The 4B and 0.6B weights, tokenizer files, model cards, and the Thinking-climate-100k dataset are on the KissanAI Hugging Face organisation. The model cards describe the architecture, intended use, training format, and known limitations.

Two documentation gaps remain. Public evaluation numbers have not been added, and the dataset licence needs to be specified. Until those are resolved, the release should be treated as a research preview rather than a production recommendation.

The most useful feedback is a concrete failure case: the question, the context the model missed, the answer it gave, and the answer a qualified reviewer would accept.