Titans-Llama-3.2-1B

Titanesque version of meta-llama/Llama-3.2-1B with parallel linearized attention (TPTT 😊) and PEFT.

The model was presented in the paper TPTT.

Model Details

• Architecture: TpttModel
• Base model: meta-llama/Llama-3.2-1B
• LiZA config: operator=delta_rule, mag=0.5
• LoRA config: r=8, alpha=16, dropout=0.05
• torch_dtype: bfloat16

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
"ffurfaro/Titans-Llama-3.2-1B",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("ffurfaro/Titans-Llama-3.2-1B")

prompt = "Your prompt here"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs, skip_special_tokens=True))

Training

• Dataset: yahma/alpaca-cleaned
• Platform: Kaggle
• Hardware: NVIDIA 2xT4
• Batch size: 3
• Epochs: 5.0
• Learning rate (final): 1.1904761904761906e-06
• Loss (final): 1.375
• Training runtime: 1654.117 sec
• Samples per second: 1.511
• Steps per second: 0.254
• Total FLOPs: 5615136276480000.0
• Gradient norm (final): 2.6798148155212402

Evaluation

• Metrics: Training loss only (no eval yet, table soon : PiQA, ARC, Hella, Wino, GSM8K, MMLU)
• Results: Final training loss: 1.375

Citation & Contact

If you use TPTT in your academic work, please cite Furfaro. For questions or support, please open an issue on the GitHub repository or contact the maintainer.

---