114 lines
3.6 KiB
Markdown
114 lines
3.6 KiB
Markdown
---
|
||
library_name: transformers
|
||
tags:
|
||
- text-generation-inference
|
||
- code
|
||
- llama-3.2
|
||
- math
|
||
- general-purpose
|
||
license: llama3.2
|
||
language:
|
||
- en
|
||
base_model:
|
||
- meta-llama/Llama-3.2-1B
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
|
||
# **Oganesson-TinyLlama-1.2B**
|
||
|
||
> **Oganesson-TinyLlama-1.2B** is a lightweight and efficient language model built on the **LLaMA 3.2 1.2B** architecture. Fine-tuned for **general-purpose inference**, **mathematical reasoning**, and **code generation**, it’s ideal for edge devices, personal assistants, and educational applications requiring a compact yet capable model.
|
||
|
||
> \[!note]
|
||
> GGUF: [https://huggingface.co/prithivMLmods/Oganesson-TinyLlama-1.2B-GGUF](https://huggingface.co/prithivMLmods/Oganesson-TinyLlama-1.2B-GGUF)
|
||
|
||
---
|
||
|
||
## **Key Features**
|
||
|
||
1. **LLaMA 3.2 1.2B Core**
|
||
Powered by the latest **TinyLLaMA (1.2B)** variant of Meta's LLaMA 3.2, offering modern instruction-following and multilingual capabilities in a very small footprint.
|
||
|
||
2. **Modular Fine-Tuning**
|
||
Trained on a handcrafted modular dataset covering general-purpose reasoning, programming problems, and mathematical challenges.
|
||
|
||
3. **Mathematical Competence**
|
||
Solves equations, explains concepts, and performs symbolic logic in algebra, geometry, and calculus—ideal for lightweight tutoring use cases.
|
||
|
||
4. **Code Understanding & Generation**
|
||
Produces clean, interpretable code in Python, JavaScript, and more. Useful for micro-agents, code assistants, and embedded development tools.
|
||
|
||
5. **Versatile Output Formats**
|
||
Handles JSON, Markdown, LaTeX, and structured data output, enabling integration into tools and platforms needing formatted results.
|
||
|
||
6. **Edge-Optimized**
|
||
At only 1.2B parameters, this model is built for **local inference**, **on-device usage**, and **battery-efficient environments**.
|
||
|
||
---
|
||
|
||
## **Quickstart with Transformers**
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_name = "prithivMLmods/Oganesson-TinyLlama-1.2B"
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
||
prompt = "Write a Python function to compute the Fibonacci sequence."
|
||
|
||
messages = [
|
||
{"role": "system", "content": "You are a helpful coding and math assistant."},
|
||
{"role": "user", "content": prompt}
|
||
]
|
||
|
||
text = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=False,
|
||
add_generation_prompt=True
|
||
)
|
||
|
||
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
|
||
generated_ids = model.generate(
|
||
**model_inputs,
|
||
max_new_tokens=512
|
||
)
|
||
generated_ids = [
|
||
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
|
||
]
|
||
|
||
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
||
print(response)
|
||
```
|
||
|
||
---
|
||
|
||
## **Intended Use**
|
||
|
||
* Lightweight reasoning for embedded and edge AI
|
||
* Basic math tutoring and symbolic computation
|
||
* Code generation and explanation for small apps
|
||
* Technical content in Markdown, JSON, and LaTeX
|
||
* Educational tools, personal agents, and low-power deployments
|
||
|
||
---
|
||
|
||
## **Limitations**
|
||
|
||
* Smaller context window than 7B+ models
|
||
* Less suitable for abstract reasoning or creative writing
|
||
* May require prompt engineering for complex technical queries
|
||
* Knowledge is limited to pretraining and fine-tuning datasets
|
||
|
||
---
|
||
|
||
## **References**
|
||
|
||
1. [LLaMA 3 Technical Report (Meta)](https://ai.meta.com/llama/)
|
||
2. [YaRN: Efficient Context Window Extension of Large Language Models](https://arxiv.org/pdf/2309.00071) |