chuanguo's picture
Update README.md
b8aba21 verified
metadata
base_model: meta-llama/Meta-Llama-3-8B-Instruct
library_name: peft
license: fair-noncommercial-research-license
datasets:
  - yahma/alpaca-cleaned
extra_gated_fields:
  First Name: text
  Last Name: text
  Date of birth: date_picker
  Country: country
  Affiliation: text
  I accept the terms and conditions: checkbox
  geo: ip_location
language:
  - en
tags:
  - facebook
  - meta
  - pytorch
  - llama
  - llama-3

TamedLlama-8B-Instruct

Repository for TamedLlama-8B-Instruct, a fine-tuned variant of Llama-3-8B-Instruct that is robust against prompt injection attacks. See our TamedLlama paper for more information.

Utility Evaluation (higher is better)

Category Benchmark Metric Llama 3 8B Instruct TamedLlama 8B Instruct GPT-4o-mini GPT-4o (2024-11-20)
General Knowledge MMLU (0-shot, CoT) macro_avg/acc 64.6 61.2 82.0[1] 85.7[2]
MMLU Pro (5-shot, CoT) macro_avg/acc 42.5 40.7 63.1[3] 77.9[3]
IFEval 76.3 74.1 - -
BBH (3-shot, CoT) acc 68.4 64.6 - -
GPQA (0-shot, CoT) acc 35.3 32.6 40.2[1] 46.0[2]
Instruction Following AlpacaEval2 win_rate 28.0 26.5 44.7 56.2
SEP win_rate 50.0 48.5 65.9 64.9

Security Evaluation (lower is better)

Category Benchmark Metric Llama 3 8B Instruct TamedLlama 8B Instruct GPT-4o-mini GPT-4o (2024-11-20)
Instruction Following AlpacaFarm ASR 23.1 0.0 0.5 0.0
SEP (start) ASR 48.7 5.9 14.6 14.8
SEP (end) ASR 60.0 6.8 9.1 14.4
TaskTracker ASR 5.3 0.2 0.3 0.6
CyberSecEval2 ASR 43.6 7.3 25.5 20.0