README.md

---
language:
- es
license: apache-2.0
base_model: unsloth/Qwen3.5-0.8B
datasets:
- FarmifAI/FarmifAI_dataset_1.3
library_name: transformers
pipeline_tag: text-generation
tags:
- agriculture
- colombia
- rag
- small-language-model
- on-device
- lora
- unsloth
- qwen3_5
---

# FarmifAI 1.3

**FarmifAI 1.3** is a small language model (fine-tuned from [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B)) that answers agricultural questions in Spanish **from a context you provide**. It is built for Colombian agriculture and powers FarmifAI, an offline assistant app for farmers: given a question and technical passages retrieved from a knowledge base, it writes a short reasoning and a clear final answer based only on that context.

This repository contains the full-precision weights. For llama.cpp and on-device use, see **[FarmifAI_1.3_GGUF](https://huggingface.co/FarmifAI/FarmifAI_1.3_GGUF)**.

> FarmifAI is not a general-purpose chatbot. It is trained to answer from the context it receives, so it should always be used together with a retrieval step.

## Model details

| | |
|---|---|
| **Developed by** | FarmifAI, Universidad del Cauca (Colombia) |
| **Base model** | Qwen3.5-0.8B |
| **Language** | Spanish |
| **Input** | System prompt + retrieved context in `<knowledge>` tags + question |
| **Output** | `<reasoning>…</reasoning>` followed by `<answer>…</answer>` |
| **License** | Apache 2.0 |

## Intended use

- Answering farmers' questions in Spanish inside a RAG pipeline, using passages from agricultural technical documents.
- Not meant to be used without retrieved context, in other languages, or as the only basis for decisions such as agrochemical selection or dosing.

## Prompt format

The model was trained with a fixed Spanish system prompt. Use it as is:

```text
Eres un asistente agrícola. Responde únicamente con la información dentro de <knowledge>. Si la respuesta no está en el contexto, declara que no tienes información; no inventes datos.

Instrucciones de formato:
- En <reasoning>, analiza paso a paso el contexto frente a la ...

<<< PASTE THE REST OF THE EXACT SYSTEM PROMPT FROM THE DATASET HERE >>>
```

The user message contains the retrieved context followed by the question:

```text
<knowledge>
{retrieved context}
</knowledge>

{question}
```

The model replies with a step-by-step analysis and a final answer:

```text
<reasoning>
Step-by-step analysis of the context against the question.
</reasoning>
<answer>
Final answer in Spanish, based only on the provided context.
</answer>
```

Applications typically show only the `<answer>` block. Parse it defensively in case the tags are missing.

## Quickstart

```python
import re
from transformers import AutoProcessor, AutoModelForMultimodalLM

model_id = "FarmifAI/FarmifAI_1.3"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(model_id, device_map="auto", dtype="auto")

SYSTEM_PROMPT = "..."  # the system prompt from "Prompt format" above

context = "..."  # passages retrieved from your knowledge base
question = "¿Cómo puedo controlar la broca en mi cultivo de café?"
user_message = f"<knowledge>\n{context}\n</knowledge>\n\n{question}"

messages = [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
    {"role": "user", "content": [{"type": "text", "text": user_message}]},
]

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.1, top_p=0.9)
text = processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)

match = re.search(r"<answer>(.*?)</answer>", text, re.DOTALL)
print(match.group(1).strip() if match else text)
```

**Recommended settings:** `temperature=0.1`, `top_p=0.9`, `max_new_tokens=512`.

## Training

Fine-tuned with LoRA (Unsloth + TRL) on [FarmifAI_dataset_1.3](https://huggingface.co/datasets/FarmifAI/FarmifAI_dataset_1.3), about 8.8k synthetic Spanish conversations built from technical documents on Colombian agriculture. Each example pairs a context passage and a question with a reasoning + answer response.

## Results

Compared with the base Qwen3.5-0.8B on 250 held-out examples, using the same prompt and settings for both models:

| Metric | Base model | FarmifAI 1.3 |
|---|---|---|
| Format adherence ↑ | 24.4% | **98.0%** |
| Answer relevancy (LLM judge, 0–5) ↑ | 2.16 | **4.70** |
| Faithfulness to context (LLM judge, 0–5) ↑ | 2.13 | **3.48** |
| Contradictions with context (NLI) ↓ | 30.4% | **17.2%** |

## Limitations

- The model can still make mistakes or add details that are not in the context. Check its recommendations, especially anything about agrochemicals, doses or safety periods.
- Answer quality depends on the quality of the retrieved context.
- It was evaluated with automatic metrics and LLM judges, not by agronomists or in the field, and it is not a substitute for professional advice. Its training material may reflect outdated practices, so do not use it for current regulations or registered products.

## Links

- Quantized versions: [FarmifAI/FarmifAI_1.3_GGUF](https://huggingface.co/FarmifAI/FarmifAI_1.3_GGUF)
- Dataset: [FarmifAI/FarmifAI_dataset_1.3](https://huggingface.co/datasets/FarmifAI/FarmifAI_dataset_1.3)

*Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth).*