Standard evaluation of Llama-3-8B on GSM8K yields 0.23% accuracy. The correct answer is present in 81.73% of responses. One format instruction recovers 71.57% — a 311x improvement. This model achieves 69.45% after Acharyulu Protocol training.

Research papers

XARL-2026-01: https://doi.org/10.5281/zenodo.19956997 (NeurIPS 2026, TMLR #10036)
XARL-2026-02: https://doi.org/10.5281/zenodo.20061129 (TMLR #10037)
XARL-2026-03: https://doi.org/10.5281/zenodo.20502323
XARL-2026-04: https://doi.org/10.5281/zenodo.20985017 (TMLR #10079)
XARL-2026-05: https://doi.org/10.5281/zenodo.21007235 (TMLR #10092)
XARL-2026-06: https://doi.org/10.5281/zenodo.21017286 (TMLR #10105)

Results

Condition	GSM8K Accuracy
Base Llama-3-8B bare	0.23%
Base Llama-3-8B instructed	71.57%
xArch Acharyulu v2 (this model)	69.45%

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "xarch-ai/xarch-llama3-8b-acharyulu-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.float16, device_map="auto"
)

messages = [{"role": "user", "content": "What is 25 x 48?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Team

Surya Sai Brahmendra Ramoju, Sri Likhitha Anuganti, Rohith Varma Mudundi, Naga Bala Venkatesh Ramoju xArch AI Research Laboratory — Northeastern University, Boston