SupraLabs/Supra2.5-Ultra-Preview-001

Supra2.1-Ultra

Supra2.1-Ultra is the model that refers to the Supra2.1 family. It has 230M params, 40B tokens trained on, with LLaMa arch. It is a small base model that is probably going to generate incorrect or biased text, and it has neither been aligned nor evaluated for safety.

Config

Spec	Details
Parameters	230M
Tokens seen	~40.4B
Arch	Llama
Layers	32
Hidden size	768
Intermediate size (MLP)	1920
Attention heads	12 (head dim 64)
KV heads (GQA)	4
Vocabulary size	49,152
Context length	2,048
RoPE theta	10,000
Activation	SiLU (SwiGLU MLP)
Tied embeddings	Yes
Precision	bfloat16

Data used to train this model

HuggingFaceFW/fineweb-edu

mlfoundations/dclm-baseline-1.0

HuggingFaceTB/finemath

Hardware

This model was trained on a single A100 SXM.

Benchmarks

Benchmark	Shots	acc	acc_norm	Random
HellaSwag	0	31.64	37.18	25.0
PIQA	0	65.78	64.91	50.0
ARC-Easy	0	57.11	50.46	25.0
ARC-Challenge	0	25.00	26.28	25.0
Arithmark-v3	0	38.00	38.20	-
Open SLM leaderboard's II	-	20.64	-	-

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "SupraLabs/Supra2.1-Ultra"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16)

prompt = "The capital of France is"
inputs = tokenizer(prompt, return_tensors="pt")

output = model.generate(
    **inputs,
    max_new_tokens=50,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))

This code should download the model and generate a response.

License

This model is released under Apache-2.0 License.

Credits

thanks to:

Dedeprogames, for code to train this model.