---
license: apache-2.0
base_model: unsloth/Llama-3.2-1B-Instruct
tags:
- safetensors
- llama-3
- merged
- tumbuka
- bantu
- malawi
language:
- tum
- en
pipeline_tag: text-generation
library_name: transformers
---

# Mazgu_Llama-1B-V2-Knowledge-16bit (Standalone Model)

**Mazgu_Llama-1B-V2-Knowledge-16bit** is a fully merged, standalone 16-bit `model.safetensors` release of Llama-3-1B, fine-tuned specifically for the Tumbuka language. 

Because the LoRA weights are merged directly into the base architecture, this repository **does not require loading external adapter weights at runtime**. It serves as an ideal baseline for:
* Direct inference via Hugging Face `transformers` or `vLLM`.
* Continual pre-training or further SFT runs.
* Full-model export and custom quantization pipelines.

---

## 💻 Quickstart Guide

### Option 1: Standard Transformers

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

model_id = "Mwanzau/Mazgu_Llama-1B-V2-Knowledge-16bit"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "### Instruction:\nLongosolani vya Yesu mu Mateyo 24.\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=150,
    temperature=0.6,
    repetition_penalty=1.2
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

### Option 2: Optimized 4-Bit Loading with Unsloth (Colab / Kaggle)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Mwanzau/Mazgu_Llama-1B-V2-Knowledge-16bit",
    max_seq_length = 2048,
    load_in_4bit = True, # Saves VRAM on T4 GPUs
)

FastLanguageModel.for_inference(model)

prompt = "### Instruction:\nLongosolani vyakurya ivyo vili bwino ku munda.\n\n### Response:\n"
inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.6, repetition_penalty=1.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

### 📊 Summary
Format: Unsharded Native safetensors
Precision: 16-bit Float (bfloat16/float16)
Vocabulary Focus: Tumbuka instruction-following and factual knowledge retrieval.