Model Card for Llama-3.1-Argunaut-1-8B-SFT

Name: Llama-3.1-Argunaut-1-8B-SFT
Availability: InStock

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct. It has been trained using TRL.

📘 HF Blog Article

Visualize in Weights & Biases

Quick start

python

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"])

Evaluation

Chat Experience


[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]

Anbother chat:


[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]

Metrics

	Llama-3.1-8B-Instruct	Argunaut-1-8B-SFT
⬇️ Argdown Bench
pass@1	80.8	98.9
pass@5	98.9	99.8
graph sim	8.4	65.5
isomorphic	1.3	44.0
🤗 HF Leaderboard
MMLU pro	37.6	34.5
MUSR	40.1	41.6
GPQA Diamond	32.8	30.8
GPQA Main	28.5	32.1
MATH	12.5	9.1
BBH	54.7	48.2
⛓️ COT Leaderboard
LogiQA	5.9	1.4
LogiQA2	15.5	0.8
LSAT-ar	11.7	3.0
LSAT-lr	20.8	3.9
LSAT-rc	27.5	13.8

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)

yaml

# Training parameters
num_train_epochs: 1
per_device_train_batch_size: 8
gradient_accumulation_steps: 2
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
learning_rate: 5.0e-6  # following _Tülu 3_ recipe
lr_scheduler_type: cosine
warmup_ratio: 0.1

Hardware: 2 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

Open LLM Leaderboard Evaluation Results

Detailed results can be found here! Summarized results can be found here!

Metric	Value (%)
Average	23.56
IFEval (0-Shot)	55.19
BBH (3-Shot)	27.19
MATH Lvl 5 (4-Shot)	11.18
GPQA (0-shot)	4.47
MuSR (0-shot)	15.85
MMLU-PRO (5-shot)	27.47