Open-sourcing Dhenu Llama 3

Within a day of Meta releasing Llama 3, we published Dhenu Llama 3, our fine-tune of Llama 3 8B for Indian agriculture. The model card is on Hugging Face at KissanAI/llama3-8b-dhenu-0.1-sft-16bit. This is the second checkpoint in the Dhenu family.

For engineers who used Dhenu 1.0, the useful questions are what changed, what held up in evaluation, and whether the migration is worth doing.

Why so fast after Llama 3

Our fine-tune pipeline does not care which supported base it points at, and the training corpus sits prepared. When the Llama 3 weights became available on the 18th, we started the run that afternoon. The quick release came from having the boring preparation done ahead of time.

The reason to be ready is that base-model improvements compound through to the domain model. A better base means a better Dhenu without us having to do the corpus work over. We pay the cost of operational readiness so the model improvements are downstream of base-layer improvements rather than gated on our own velocity.

What changed from Dhenu 1.0

Dhenu 1.0 was built on Sarvam’s OpenHathi (a Llama-2-7B Indic adaptation), released December 2023. Dhenu Llama 3 is built on Llama 3 8B base, released this week. The concrete improvements:

Multilingual fluency. Llama 3’s tokenizer and pretraining handle Indic scripts substantially better than the Llama 2 family. The Devanagari handling alone produces noticeably cleaner outputs. Hindi, Marathi, and Gujarati responses from Dhenu Llama 3 are visibly more fluent than the Dhenu 1.0 outputs on the same prompts. Telugu, Tamil, and Bengali show the same improvement.

Reasoning depth. Llama 3 8B is a stronger reasoner than the OpenHathi base, and that capability propagates through the fine-tune. The model can hold multi-conditional agronomic context better. A query that involves “given the rainfall, the growth stage, and the regional pest history, should I apply X” is handled more coherently in Dhenu Llama 3 than in Dhenu 1.0.

Eval scores. On our internal review set (the same rubric we used for Dhenu 1.0, plus 1,200 questions added since), Dhenu Llama 3 reads clearly ahead of Dhenu 1.0, with the improvement concentrated in the multi-conditional questions, which is the case we care about most. The usual caveat: this is our own rubric reviewed by our own team, and it will stay that way until a public benchmark for agricultural AI exists.

What it is now usable for

The same use cases we have always cared about, but with a higher quality floor.

A farmer-facing voice assistant in one of the major Indian languages, integrated into an agribusiness’s distribution channel through our deployment stack. The voice path goes through the same speech recognition and synthesis layers we have built; the language model is the new component. For teams experimenting with Dhenu 1.0, the migration path is straightforward.

An agronomy retrieval assistant for an extension officer or a retailer staff member. This is the use case where the reasoning depth matters most: the person asking the question often has more context than a smallholder farmer, and the model needs to handle multi-turn refinement of a recommendation.

A research / prototyping base for other teams building agricultural AI. The model is on Hugging Face under a permissive licence (Llama 3 Community Licence, which we are passing through). Fine-tuning Dhenu Llama 3 further for a specific crop, region, or business is straightforward and is the kind of thing we explicitly want to encourage.

Training details

The fine-tune is full supervised fine-tuning, not LoRA. We ran an adapter-only pass first; the gains on multi-conditional questions did not hold up against full SFT on the 8B base, so we paid for the full run. The training set is the same Dhenu agricultural corpus we have built up, now an instruction corpus several times larger than the 300,000 pairs Dhenu 1.0 trained on. The full SFT run completed within the day. We are keeping the claim at that level because the exact duration depends on which preparation and validation steps are counted as part of the run.

We released the weights at 16-bit precision. Quantised versions for edge deployment will follow. The model card details the training config, the prompt template, and the evaluation harness.

Why open

The reasoning has not changed since December. Open weights mean the gaps in the model are inspectable, and the next team can build on what worked instead of reverse-engineering it. What is hard to copy about KissanAI is the corpus curation built up over years and the deployment layer around the models; neither ships in a checkpoint. Public weights also work as a forcing function on quality, because other researchers find what we missed faster than our own testing does.

What is next

The next checkpoint in the Dhenu line is underway. The recurring theme in the current round of corpus work is climate stress: erratic rainfall and heat show up in incoming farmer queries more often than the original corpus would predict, and the curation is following the queries. On the vision side, Dhenu Vision shipped last month and continues as its own line.

The Dhenu Llama 3 weights are on Hugging Face. We are working on the quantised edge builds next.