abrhaleyarefaine1997/gpt2-tigrinya-lora

GPT-2 Tigrinya LoRA Fine-Tuned Model --- base_model: gpt2 library_name: peft pipeline_tag: text-generation tags: - base_model:adapter:gpt2 - lora - transformers datasets: - custom metrics: - perplexity language: - ti new_version: 1.0 --- # Low-Resource Tigrinya Language Modeling with LoRA-Fine-Tuned GPT-2 ## Model Details ### Model Description This model is a **GPT-2 small** (124M parameters) fine-tuned using **LoRA (Low-Rank Adaptation)** on a custom Tigrinya dataset. It is designed for **Tigrinya text generation**, including chatbot/dialogue, storytelling, and text continuation. - **Developed by:** Abrhaley (Warsaw University of Technology, MSc student) - **Funded by [optional]:** N/A - **Shared by [optional]:** Abrhaley - **Model type:** Causal Language Model (decoder-only Transformer) - **Language(s) (NLP):** Tigrinya (`ti`) - **License:** MIT - **Finetuned from model [optional]:** gpt2 - **Framework versions:** Transformers + PEFT 0.17.1 ### Model Sources - **Repository:** GitHub - **Paper [optional]:** N/A - **Demo [optional]:** N/A --- ## Uses ### Direct Use - Tigrinya text generation - Chatbot / dialogue systems - Story or content generation ### Downstream Use [optional] - Fine-tuning for domain-specific Tigrinya applications (news, education, cultural storytelling) ### Out-of-Scope Use - Generating harmful, offensive, or misleading content - Critical decision-making without human supervision --- ## Bias, Risks, and Limitations - Dataset may not cover all Tigrinya dialects - May generate biased, offensive, or incoherent outputs - Not suitable for factual QA tasks ### Recommendations Users should verify outputs before real-world use and avoid sensitive applications. --- ## How to Get Started with the Model ```python from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline model_id = "abrhaley/gpt2-tigrinya-lora" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id) generator = pipeline("text-generation", …