|
|
--- |
|
|
library_name: transformers |
|
|
tags: |
|
|
- unsloth |
|
|
- lora |
|
|
- gemma |
|
|
- whatsapp |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- es |
|
|
base_model: |
|
|
- unsloth/gemma-2-2b-it-bnb-4bit |
|
|
--- |
|
|
|
|
|
# Model Card for Whatsapp-Finetuned LoRA (Gemma-2-2B-IT-4bit) |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
This is a LoRA adapter trained on personal WhatsApp conversations, applied on top of **`unsloth/gemma-2-2b-it-bnb-4bit`**, an instruction-tuned Gemma 2B model in 4-bit quantization. |
|
|
The adapter specializes the base model toward informal Spanish conversational style, slang, and context typical of WhatsApp chats. |
|
|
|
|
|
- **Developed by:** Private (Boni) |
|
|
- **Model type:** LoRA adapter for causal language modeling |
|
|
- **Language(s):** Spanish (es) |
|
|
- **Finetuned from model:** `unsloth/gemma-2-2b-it-bnb-4bit` |
|
|
|
|
|
### Model Sources |
|
|
- **Base model:** [unsloth/gemma-2-2b-it-bnb-4bit](https://huggingface.co/unsloth/gemma-2-2b-it-bnb-4bit) |
|
|
|
|
|
--- |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
- Chatbots and assistants that mimic WhatsApp-style Spanish conversations. |
|
|
- Experimentation with low-rank adapters on personal datasets. |
|
|
|
|
|
### Downstream Use |
|
|
- Can be merged with the base model for full fine-tuned inference. |
|
|
- Can be combined with other adapters for multi-domain behavior. |
|
|
|
|
|
### Out-of-Scope Use |
|
|
- Production deployment without careful filtering (the dataset is personal, informal, and may not generalize). |
|
|
- Sensitive domains like healthcare, law, or safety-critical applications. |
|
|
|
|
|
--- |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
- The dataset consists of personal WhatsApp conversations, which may include biases, informal expressions, and idiosyncratic slang. |
|
|
- The model may reflect private communication style and does not guarantee factual correctness. |
|
|
- Limited training size means performance outside the conversational domain is reduced. |
|
|
|
|
|
### Recommendations |
|
|
Users should treat outputs as experimental. Avoid relying on this model for factual or professional contexts. |
|
|
|
|
|
--- |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
from peft import PeftModel |
|
|
|
|
|
base_model = "unsloth/gemma-2-2b-it-bnb-4bit" |
|
|
adapter = "Jabr7/Mini-Boni" |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(base_model) |
|
|
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto") |
|
|
model = PeftModel.from_pretrained(model, adapter) |
|
|
|
|
|
prompt = "Hola, ¿cómo estás?" |
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda") |
|
|
outputs = model.generate(**inputs, max_new_tokens=100) |
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
|
