---
language:
- en
- tl
- ceb
- war
- hil
- pam
- bcl
- ilo
thumbnail: https://res.cloudinary.com/digpzthsf/image/upload/v1738737064/grdtuk20lbskymaenvvt.jpg
tags:
- deepseek
- distill
- gossip
- philippines
- multi-language
- unsloth
- llm
- chat
library_name: peft
license: mit
datasets:
- jhenberthf/filipino-gossip-dataset
metrics:
- accuracy
- perplexity
base_model: unsloth/DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit
---

# Marites AI 🗣️🔥

**_A Text Generation Model for Crafting Humorous Gossip (Chismis) with a Filipino Flair_** 🗣️🔥

![Marites AI](https://res.cloudinary.com/digpzthsf/image/upload/v1738734237/jdxmcwhpnv4ix7mcjl5t.jpg)

## Model Description

**Marites AI** is a playful text generation model designed to create gossip ("chismis") infused with humor and wit, mimicking the style of real-life gossip mongers (inspired by the iconic Filipino "Marites" persona). Whether you want to generate lighthearted drama, funny rumors, or exaggerated scenarios, this model serves up entertaining and culturally nuanced chismis perfect for memes, creative writing, or laughter therapy.

This model is a **PEFT/LoRA adapter** fine-tuned from the **DeepSeek-R1-Distill-Llama-8B** base model (an 8B-parameter Llama-style model, loaded in 4-bit via `unsloth/DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit`) by DeepSeek AI, then specialized for Filipino/Taglish gossip generation on the [`jhenberthf/filipino-gossip-dataset`](https://huggingface.co/datasets/jhenberthf/filipino-gossip-dataset) (v1.2 — 272 rows across 8 Philippine languages).

### Key Features

- **Humor-Packed Gossip**: Generates absurd, over-the-top, and sarcastic chismis.
- **Cultural Nuance**: Optimized for Filipino/Taglish (Tagalog-English) slang and contexts.
- **Customizable Drama**: Control the "spiciness" of gossip with adjustable parameters.
- **Built on Transformers**: Leverages a state-of-the-art language model architecture.

## Training Data

Marites AI was fine-tuned on the [`jhenberthf/filipino-gossip-dataset`](https://huggingface.co/datasets/jhenberthf/filipino-gossip-dataset) — a collection of Filipino/Taglish gossip prompt–response pairs in multiple Philippine languages and dialects.

- **Version:** v1.2 (272 rows: 263 train / 9 test)
- **Languages:** Tagalog, Bisaya, Ilocano, Hiligaynon, Waray, Kapampangan, Bikol, and English code-switching
- **Schema:** `prompt`, `response`, `category`, `tags`, `persona`
- **Style:** each row is voiced by a fictional "Marites" persona (e.g. *Doña Chismosa*, *Barangay Tambay*) — satirical and clearly fictional, never about real individuals
- **Enhancement:** expanded from 47 → 272 rows; all text passes through a cleaning pipeline (mojibake fix, dedup, tag normalization) before training

## Installation

```bash
pip install transformers torch peft bitsandbytes accelerate
```

## Usage

### Quick start (GPU required)

Marites AI is a LoRA adapter on top of an 8B 4-bit base, so it must be loaded on a **GPU** (bitsandbytes). You cannot load it on CPU-only hardware.

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from peft import PeftModel

BASE_ID = "unsloth/DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit"
ADAPTER_ID = "jhenberthf/marites-ai"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER_ID)
base = AutoModelForCausalLM.from_pretrained(BASE_ID, device_map="auto")
model = PeftModel.from_pretrained(base, ADAPTER_ID)
marites = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)

prompt = "Alam mo ba si Maria at Juan..."
chismis = marites(
    prompt,
    max_new_tokens=100,
    temperature=1.2,   # Higher = more chaotic
    repetition_penalty=1.5,
)
print(chismis[0]["generated_text"])
```

#### Example Output

**_"Alam mo ba si Maria at Juan? Nakita ko silang nagtatanan sa CR ng Jollibee! Tapos may dala-dalang dalawang burger steak, sabi raw 'para di magutom habang nagkakalatayan ng sama ng loob kay Tita Glenda!' Chz, pero totoo yung nakita kong resibo sa bag ni Maria—₱500 pala utang niya kay Juan. Aba'y love loan!"_**

#### Advance Parameters

- **temperature:** Increase (1.5+) for wilder stories, decrease (0.7-) for more coherence.
- **max_length:** Keep it between 50-150 for optimal drama-to-readability ratio.
- **repetition_penalty:** Use 1.2-2.0 to minimize repetitive plot twists.

## Base Model

Marites AI is a PEFT/LoRA adapter fine-tuned from **DeepSeek-R1-Distill-Llama-8B** — an 8B-parameter Llama-style model. The adapter weights were trained on top of the 4-bit (`bnb-4bit`) base. Because the underlying model is 8B, inference requires a GPU with enough VRAM for the 4-bit base (~6–8 GB) and the `bitsandbytes` library; it will not run on CPU-only environments.

## Ethical Considerations

**Use Responsibly!** 🚨

- This model generates fictional content. Do not use it to spread misinformation.
- Outputs may reflect biases in the training data. Always add disclaimers when sharing.
- Intended for entertainment only. Respect privacy and avoid targeting real individuals.

## Limitations

- **Not a Factual Source:** All chismis are AI-generated fiction.
- **Cultural Specificity:** Best results require familiarity with Filipino/Taglish humor.
- **Over-the-Top Outputs:** May generate nonsensical or exaggerated scenarios (by design!).
- **Hardware Requirement:** Needs a GPU to load the 8B 4-bit base (no CPU-only support).

## Contributing

Found a bug or want to improve Marites' chismis skills?

Open an issue or PR on [GitHub](https://github.com/jhenbertgit/marites-ai/pulls).

Share your funniest outputs in the "Community Examples" discussion tab!

## License

This model is licensed under the MIT License.

---

**Disclaimer:** This project is for entertainment purposes only. The name "Marites" refers to a cultural stereotype and is not affiliated with any real person. Generated content does not reflect the views of the model creators.

---

**Let's Keep the Chismis Alive!** 🌟
**_Star this repo, share your spiciest outputs with #MaritesAI, and tag us!_**