---
base_model: openbmb/MiniCPM4-8B
library_name: peft
pipeline_tag: text-generation
tags:
- twi
- akan
- ghana
- lora
- peft
- llama-factory
license: apache-2.0
language:
- tw
- en
---

# TwiGen MiniCPM4

LoRA adapter for [MiniCPM4-8B](https://huggingface.co/openbmb/MiniCPM4-8B) fine-tuned on Twi (Akan) instruction data. Part of the TwiGen project.

Generates coherent Twi text from English instructions.

## Usage

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

base_id = "openbmb/MiniCPM4-8B"
lora_id = "ghananlpcommunity/twi-gen-minicpm4"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    base_id, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="cuda:0",
)
model = PeftModel.from_pretrained(model, lora_id)
model.eval()

messages = [{"role": "user", "content": "Write a story about a farmer in Ghana."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False) + "<|im_start|>assistant\n"

inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=2048).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

## Training

- **Base model:** MiniCPM4-8B
- **Data:** ~495k Twi instruction samples (translated + synthetic)
- **Method:** LoRA (rank 16, alpha 32)
- **Hardware:** NVIDIA H200 (140 GB VRAM)
- **Framework:** LLaMA-Factory

## Model

| Model | Adapter |
|-------|---------|
| [twi-gen-qwen](https://huggingface.co/ghananlpcommunity/twi-gen-qwen) | Qwen3.5-0.8B version (lighter, faster) |