HachiML/Llama-2-13b-hf-qlora-dolly-ja-2ep

JGLUE Score

I evaluated this model using the following JGLUE tasks. Here are the scores:

Task	Llama-2-13b-hf(*)	This Model
JCOMMONSENSEQA(acc)	75.06	75.78
JNLI(acc)	22.18	50.69
MARC_JA(acc)	38.83	79.64
JSQUAD(exact_match)	76.13	62.83
Average	53.05	67.23

• Note: Use v0.3 prompt template
• The JGLUE scores were measured using the following script: Stability-AI/lm-evaluation-harness

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer
from peft import PeftModel

model_name = "meta-llama/Llama-2-13b-hf"
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
pt_model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
)

peft_name = "HachiML/Llama-2-13b-hf-qlora-dolly-ja-2ep"
model = PeftModel.from_pretrained(
    pt_model,
    peft_name,
)

Training procedure

The following bitsandbytes quantization config was used during training:

• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: float16

Framework versions

• PEFT 0.4.0