metadata

library_name: transformers
language:
  - la
  - en
datasets:
  - hathibelagal/clean_latin
base_model:
  - meta-llama/Llama-3.2-3B

Model Card for Llama-3.2-Latin

hathibelagal/llama-3.2-latin is a finetuned version of the LLaMA-3.2-3B model, optimized for generating and understanding Latin text across various historical periods, from ancient to modern Neo-Latin. At this point, it generates content with accurate use of tenses (e.g., pluperfect, subjunctive), cases, and complex structures (e.g., concessive, temporal clauses).

Intended Use

Suitable for tasks involving Latin text generation, translation, or analysis, such as generating Classical Latin prose, completing sentences, or aiding in Latin education. Best for short, context-specific prompts due to coherence limitations.
Not recommended for long-form narrative generation or tasks requiring strict contextual consistency until coherence improves.

Ethical Considerations

Bias: The model may reflect biases in the training dataset, such as overrepresentation of certain Latin styles (e.g., ecclesiastical Latin) leading to tonal shifts in output.
Usage: Generated text should be reviewed for accuracy, especially in educational or scholarly contexts.

Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo_id = "hathibelagal/llama-3.2-latin"

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

prompt = "Libellus vere aureus"
inputs = tokenizer.encode(prompt, return_tensors="pt")
inputs = inputs.to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=100,
    do_sample=True,
    top_p=0.9,
    temperature=0.6,
    repetition_penalty=1.2
)

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