Model Card for Llama-3.3-Argunaut-1-70B-SFT

This model is a fine-tuned version of meta-llama/Llama-3.3-70B-Instruct. It has been trained using TRL.

Quick start

from transformers import pipeline

question = "Are you familiar with Argdown syntax? What's its purpose?"
generator = pipeline("text-generation", model="DebateLabKIT/Llama-3.1-Argunaut-1-8B-SFT", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

SFT dataset mixture

Dataset	Weight (examples)	Weight (tokens)
DebateLabKIT/deepa2-conversations	25%	49%
DebateLabKIT/deep-argmap-conversations	25%	18%
allenai/tulu-3-sft-mixture	50%	33%

Training procedure

Trained with SFT on 1M examples and for 1 epoch with

• context length 8196
• packing (trl implementation)
• spectrum (top 30 percent)

# Training parameters
num_train_epochs: 1
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
learning_rate: 2.0e-6  # following _Tülu 3_ recipe
lr_scheduler_type: cosine
warmup_ratio: 0.1

Hardware: 4 x H100 GPUs.

This work was performed on the HoreKa supercomputer funded by the Ministry of Science, Research and the Arts Baden-Württemberg and by the Federal Ministry of Education and Research.

Framework versions

• TRL: 0.12.1
• Transformers: 4.46.3
• Pytorch: 2.4.1
• Datasets: 3.1.0
• Tokenizers: 0.20.3

Credits

This work wouldn't be possible without all the great contributions from the open LLM community. Thank you! Special kudos go to

• @philschmid for his latest fine-tuning boilerplate
• @lvwerra, @lewtun et al for building and maintaining trl
• @cognitivecomputations for sharing spectrum
• @allenai for the Tülu recipe and artifacts