GeoV-9b-r2

GeoV-9B-r2 is a 9 billion parameter causal language model.

It is still being trained and has the same architecture as the GeoV-9b model, but the training data is sampled without replacement; (GeoV-9b models training data was sampled with replacement).

The GeoV model was designed by Georges Harik and uses Rotary Positional Embeddings with Relative distances (RoPER) by Georges Harik and Varuna Jayasiri.

RoPER, in addition to using relative positions in the attention score calculation by RoPE embeddings, adds relative positional information explicitly to value embeddings. Specifically, it incorporates the relative positions of the tokens paid attention to. RoPER has given better performance in some algorithmic tasks, and seems comparable to RoPE in language modeling.

Model details

• Developed by: Georges Harik
• Model type: Transformer-based Language Model
• Language: English

<figure style="width:30em"> </figure>

The current released weights were trained on ~39 billion tokens. We plan to continue training up to 300 billion tokens. This training run is monolingual and uses c4en and english wikipedia datasets.

Test results

These are the results from EleutherAI/lm-evaluation-harness at 81B (tokens trained) checkpoint.

Installation

pip install geov

Generation

from geov import GeoVForCausalLM, GeoVTokenizer

model = GeoVForCausalLM.from_pretrained("GeoV/GeoV-9b-r2")
tokenizer = GeoVTokenizer.from_pretrained("GeoV/GeoV-9b-r2")

prompt = "In mathematics, topology is the study of"

input_ids = tokenizer(prompt, return_tensors="pt").input_ids

gen_tokens = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.9,
    max_length=100,
)
gen_text = tokenizer.batch_decode(gen_tokens)[0]