Text Generation
Transformers
PyTorch
Japanese
llama
text-generation-inference
Inference Endpoints

About

This model is Lightblue's QLoRA finetune of OpenOrca's Open-Orca/OpenOrcaxOpenChat-Preview2-13B model on Japanese fine-tuning datasets.

This model specialises on answering Closed Question Answering in Japanese. Input a piece of reference text, ask a question, and see the model answer based on the reference text.

We trained on equal samples of the following three datasets:

which resulted in a dataset of 13,167 samples total.

These three datasets were chosen as they represent three distinct fine-tuning tasks (Text simplification, question answering, and text summarization, respectively) which we hypothesize can help to improve the language models suitability for dealing with Japanese data. These three datasets make up the model name: STX.

With these datasets, we achieve the following scores on the JGLUE benchmark:

Model Name Open-Orca/OpenOrcaxOpenChat-Preview2-13B lightblue/openorca_stx
jsquad-1.1-0.3 0.692 0.836
jcommonsenseqa-1.1-0.3 0.831 0.782
jnli-1.1-0.3 0.504 0.48
marc_ja-1.1-0.3 0.936 0.959

We achieved these scores by using the lm-evaluation-harness from Stability AI using the below commands:

MODEL_ARGS=pretrained=lightblue/openorca_stx,use_accelerate=True
TASK="jsquad-1.1-0.3,jcommonsenseqa-1.1-0.3,jnli-1.1-0.3,marc_ja-1.1-0.3"
export JGLUE_OUTPUT_DIR=../jglue_results/$MODEL_NAME/$DATSET_NAME/$DATASET_SIZE
mkdir -p $JGLUE_OUTPUT_DIR
python main.py --model hf-causal-experimental --model_args $MODEL_ARGS --tasks $TASK --num_fewshot "2,3,3,3" --device "cuda" --output_path $JGLUE_OUTPUT_DIR/result.json --batch_size 4 > $JGLUE_OUTPUT_DIR/harness.out 2> $JGLUE_OUTPUT_DIR/harness.err

Our model achieves much better results on the question answering benchmark (JSQuAD) than the base checkpoint without monstrous degradation of performance on multi-choice question benchmarks (JCommonSense, JNLI, MARC-Ja) purely through QLoRA training. This shows the potential for applying strong language models such as Open-Orca/OpenOrcaxOpenChat-Preview2-13B to minimal QLoRA fine-tuning using Japanese fine-tuning datasets to achieve better results at narrow NLP tasks.

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

model_dir = "lightblue/openorca_stx"

tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForCausalLM.from_pretrained(
    model_dir, torch_dtype=torch.bfloat16, device_map='auto',
)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

def do_closed_qa(context, question):
    return context + "\n\n" + question

test_article = """ใ€€ใƒขใƒŽใƒžใƒใฎใƒฌใƒ‘ใƒผใƒˆใƒชใƒผใซใ€Œใƒชใƒผใƒใƒปใƒžใ‚คใ‚ฑใƒซ้ธๆ‰‹ใ€ใŒใ‚ใ‚‹ใƒฌใ‚คใ‚ถใƒผใƒฉใƒขใƒณRGใ•ใ‚“ใ€‚ๆœฌไบบๅ…ฌ่ชใฎใƒขใƒŽใƒžใƒใงใ™ใŒใ€ใƒฉใ‚ฐใƒ“ใƒผใƒ•ใ‚กใƒณใฎๅๅฟœใซๅฐ‘ใ—้ฉšใ„ใŸใใ†ใงใ™ใ€‚
ใ€€ใƒชใƒผใƒใƒปใƒžใ‚คใ‚ฑใƒซ้ธๆ‰‹ใฎใƒขใƒŽใƒžใƒใฏใ€ไฝ•ใŒใใฃใ‹ใ‘ใงใ™ใ‹ใ€‚
ใ€Œ2015ๅนดใฎใƒฏใƒผใƒซใƒ‰ใ‚ซใƒƒใƒ—๏ผˆWๆฏ๏ผ‰ใ‚คใƒณใ‚ฐใƒฉใƒณใƒ‰ๅคงไผšใงๆ—ฅๆœฌใŒๅ—ใ‚ขใƒ•ใƒชใ‚ซใ‚’ๅ€’ใ—ใŸๆฌกใฎๆ—ฅใŒใ€ไบฌ้ƒฝใงใฎ็•ช็ต„ใƒญใ‚ฑใงใ—ใŸใ€‚ๅฝ“ๆ™‚ใฏใ€ใ‚ขใƒƒใƒ—ใƒซใฎๅ…ฑๅŒๅ‰ตๆฅญ่€…ใ‚นใƒ†ใ‚ฃใƒผใƒ–ใƒปใ‚ธใƒงใƒ–ใ‚บใฎใƒขใƒŽใƒžใƒใฐใ‹ใ‚Šใงใ—ใŸใŒใ€ไธ€็ท’ใซใƒญใ‚ฑใ‚’ใ—ใฆใ„ใŸใ‚ธใƒฃใƒณใ‚ฐใƒซใƒใ‚ฑใƒƒใƒˆใ‹ใ‚‰ใ€Žใƒชใƒผใƒใƒปใƒžใ‚คใ‚ฑใƒซใซไผผใฆใพใ™ใ‚ˆใ€‚ใ‚ธใƒงใƒ–ใ‚บใฎใพใพใ€ใ„ใ‘ใ‚‹ใ‚“ใ˜ใ‚ƒใชใ„ใงใ™ใ‹๏ผŸใ€ใจ่จ€ใ‚ใ‚ŒใŸใฎใŒๅง‹ใพใ‚Šใงใ™ใ€
ใ€ŒใŸใ ใ€ใฟใ‚“ใช็Ÿฅ่ญ˜ใŒใชใ„ใ€‚ใƒฉใ‚ฐใƒ“ใƒผใ‚ทใƒงใƒƒใƒ—ใ‚’ๆŽขใ—ใ€ๆ—ฅๆœฌไปฃ่กจใฎใƒฆใƒ‹ใƒ›ใƒผใƒ ใŒๅฃฒใ‚Šๅˆ‡ใ‚Œใ ใฃใŸใฎใงใ€่ตคใฃใฝใ„ใƒฆใƒ‹ใƒ›ใƒผใƒ ใจใƒ”ใƒใƒ”ใƒใฎ็Ÿญใƒ‘ใƒณใ‚’ใฏใ„ใฆใ€‚ใจใ‚Šใ‚ใˆใšSNSใงใ€Žใƒชใƒผใƒใƒปใƒžใ‚คใ‚ฑใƒซใงใ™ใ€ใฃใฆใ„ใฃใฑใ„ๅ†™็œŸใ‚’่ผ‰ใ›ใพใ—ใŸใ€
ใ€Œใ™ใ‚‹ใจใ€ใใ‚Œใ‚’่ฆ‹ใŸใƒชใƒผใƒใ•ใ‚“ๆœฌไบบใ‹ใ‚‰DM๏ผˆใƒ€ใ‚คใƒฌใ‚ฏใƒˆใƒกใƒƒใ‚ปใƒผใ‚ธ๏ผ‰ใŒๅฑŠใใพใ—ใŸใ€‚ใ€ŽใƒขใƒŽใƒžใƒใ‚ใ‚ŠใŒใจใ†ใ”ใ–ใ„ใพใ™ใ€‚ใ‚‚ใ—ใƒขใƒŽใƒžใƒใ‚’ใ™ใ‚‹ใชใ‚‰ใ€ๅƒ•ใฎใƒฆใƒ‹ใƒ›ใƒผใƒ ใ‚’้€ใ‚Šใพใ™ใฎใง็€ใฆใใ ใ•ใ„ใ€ใจใ€‚WๆฏๅพŒใซใƒฆใƒ‹ใƒ›ใƒผใƒ 2็€ใจใƒ‘ใƒณใƒ„ใ‚„ใ‚ฝใƒƒใ‚ฏใ‚นใชใฉใ‚’ใปใ‚“ใพใซ้€ใฃใฆใใฆใใ‚Œใพใ—ใŸใ€‚ไปŠ็€ใฆใ„ใ‚‹ใฎใŒใใ‚Œใงใ™ใ€
ใ“ใ‚Œใพใงใ€ๆ•ฐใ€…ใฎ่‘—ๅไบบใ‚’ใƒขใƒŽใƒžใƒใ—ใฆใ“ใ‚‰ใ‚Œใพใ—ใŸใ€‚ใƒชใƒผใƒ้ธๆ‰‹ใฎใƒใ‚ฟใฎๅ้Ÿฟใฏใ„ใ‹ใŒใงใ—ใŸใ‹ใ€‚
ใ€€ใ€Œๅƒ•ใฏใƒฉใ‚ฐใƒ“ใƒผ็ตŒ้จ“ใŒใชใ„ใงใ™ใ—ใ€ใƒฉใ‚ฐใƒ“ใƒผใ‚’ๅ…จ็„ถ็Ÿฅใ‚‰ใชใ‹ใฃใŸใ‘ใฉใ€ใ‚„ใฃใฑใ‚Šๆœฌไบบใ‹ใ‚‰ใƒฆใƒ‹ใƒ›ใƒผใƒ ใ‚’้ ‚ใ„ใฆใ‚‹ใฃใฆใ„ใ†โ€œๅฐ็ฑ ๏ผˆใ„ใ‚“ใ‚ใ†๏ผ‰โ€ใฟใŸใ„ใชใฎใŒใ‚ใฃใฆใ€‚ใ€Žใ‚ใ„ใคใฏใƒชใƒผใƒใ•ใ‚“ๆœฌไบบใซ่ชใ‚ใ‚‰ใ‚Œใฆใ‚‹ใ€ใจใ€‚ไธ€็›ฎ็ฝฎใ‹ใ‚Œใฆใ„ใ‚‹ใฎใ‹ใชใจๆ„Ÿใ˜ใพใ™ใ€
ใ€€ใ€Œใ‚„ใฃใฆใ„ใ‚‹ใ“ใจใฏใ€่ฆ‹ใŸ็›ฎใ‚’ๆœฌไบบใซๅฏ„ใ›ใฆใƒฏใƒณใƒใƒผใƒ ใฃใฆ่จ€ใ†ใ ใ‘ใชใ‚“ใงใ™ใ‘ใฉใญใ€‚ใใ‚Œใงใ‚‚ใ€Žใ‚ใ‚ใ€ใƒชใƒผใƒใ•ใ‚“ใ ใ€ใจ่จ€ใฃใฆใ‚‚ใ‚‰ใˆใพใ™ใ€
ใ€€ใ€Œใƒชใƒผใƒใ•ใ‚“ใจๅฎŸ้š›ใซไผšใ†ใ“ใจใชใ‚“ใฆใ€็ฐกๅ˜ใซใฏใงใใชใ„ใ˜ใ‚ƒใชใ„ใงใ™ใ‹ใ€‚ใงใ‚‚ใ€ใƒชใƒผใƒใ•ใ‚“ใฎใพใญใ‚’ใ—ใฆใ„ใ‚‹RGใซใฏไผšใˆใŸใ‚ใ€ใฟใŸใ„ใช๏ผˆ็ฌ‘๏ผ‰ใ€‚ไฝ•ใ ใ‚ใ†ใชใ€ๆœ‰ๅใช็ฅž็คพใฎๆ”ฏ็คพใฎใ‚ˆใ†ใชๅญ˜ๅœจใงใ™ใ‹ใญใ€‚ใ‚ใ‚ŠใŒใŸใŒใ‚‰ใ‚Œใ‚‹ใจใ„ใ†ๆ„ๅ‘ณใงใฏไป–ใฎใƒขใƒŽใƒžใƒใจใฏใ™ใ”ใ้•ใ„ใพใ™ใญใ€
"""

test_question = "ใ€€ใƒชใƒผใƒใƒปใƒžใ‚คใ‚ฑใƒซใฏไฝ•ใ‚’้€ใฃใฆใใพใ—ใŸใ‹๏ผŸ"

pipe(do_closed_qa(test_article, question), max_new_tokens=128, temperature=0)[0]["generated_text"]
# "ใƒฆใƒ‹ใƒ›ใƒผใƒ 2็€ใจใƒ‘ใƒณใƒ„ใ‚„ใ‚ฝใƒƒใ‚ฏใ‚นใชใฉ"

Prompting

We have found that this model is able to work well using a variety of prompts, including the Alpaca style templated prompts:


f"""
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Input:
{input}
### Response:
"""

We have found that having a newline at the end of the prompt can be important for signalling that the model must respond and not continue the inputs.

Training details

We trained using the following three minimalistic prompt templates for the three tasks in STX:

  • SNOW
    f"""ๅ…ƒใฎๆ—ฅๆœฌ่ชž๏ผš
    {original_ja}
    
    ใ‚ทใƒณใƒ—ใƒซใชๆ—ฅๆœฌ่ชž๏ผš"""
    
  • TyDiQA
    f"""{passage_text}
    
    {question_text}"""
    
  • XLSum
    f"""่จ˜ไบ‹๏ผš
    {article_text}
    
    ่ฆ็ด„๏ผš"""
    

This model was trained for 1000 steps (1.2 epochs) with the model being evaluated every 50 steps. We then chose the best model from these evaluations based on validation loss. We used the qlora package from artidoro. We trained with the following hyperparameters:

Per device evaluation batch size: 16
Per device train batch size: 8
LoRA (lora_r): 64
LoRA alpha (lora_alpha): 16
LoRA modules: all
Double quantization: Enabled
Quantization type: nf4
BF16: Enabled
Bits: 4
Warmup ratio: 0.03
Learning rate scheduler type: Constant
Gradient checkpointing: Enabled
Gradient accumulation steps: 2
Learning rate: 0.0002
Adam beta2: 0.999
Maximum gradient norm: 0.3
LoRA dropout: 0.05
Weight decay: 0.0

image/png

image/png

Downloads last month
20
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for lightblue/openorca_stx

Quantizations
3 models

Datasets used to train lightblue/openorca_stx