Update README.md

README.md

@@ -1,5 +1,114 @@
----
-license: other
-license_name: tongyi-qianwen-license
-license_link: LICENSE
----
+---
+license: other
+license_name: tongyi-qianwen-license
+license_link: LICENSE
+language:
+- en
+- ja
+library_name: transformers
+pipeline_tag: text-generation
+---
+
+# nekomata-7b-pfn-qfin
+
+## Model Description
+nekomata-7b-pfn-qfin is a fine-tuned model based on [rinna/nekomata-7b](https://huggingface.co/rinna/nekomata-7b/tree/main).
+This is the base model, which is good at generating continuous sentences for finance.
+nekomata-7b-pfn-qfin is fine-tuned on 370M tokens from multiple special datasets generated by Preferred Networks, which is clear to use for commercial usage.
+The fine-tuned were carried out at a 2048 context length.
+This model is released under [Tongyi Qianwen LICENSE AGREEMENT](https://github.com/QwenLM/Qwen/blob/e8e15962d897714944773cca57fa2e460a3655e8/Tongyi%20Qianwen%20LICENSE%20AGREEMENT).
+
+The research article is available on [arXiv](https://arxiv.org/abs/2404.10555).
+
+# Benchmarking
+The benchmark score is obtained using [Japanese Language Model Financial Evaluation Harness](https://github.com/pfnet-research/japanese-lm-fin-harness)
+For the benchmark, 0-shot and default prompts are used.
+```
+|      Task      |Metric|  nekomaba-7b   |       Ours      |
+|----------------|------|------|---|------|------|---|------|
+|chabsa          |f1    |0.8134|   |      |0.8127|   |      |
+|cma_basics      |acc   |0.3158|±  |0.0764|0.3684|±  |0.0793|
+|cpa_audit       |acc   |0.2085|±  |0.0203|0.1809|±  |0.0193|
+|fp2             |acc   |0.2484|±  |0.0198|0.2674|±  |0.0203|
+|security_sales_1|acc   |0.4912|±  |0.0668|0.5088|±  |0.0668|
+|----------------|------|------|---|------|------|---|------|
+|OVER ALL        |      |0.4155           |0.4276           |
+```
+## Usage
+Install the required libraries as follows:
+```sh
+>>> python -m pip install numpy sentencepiece torch transformers accelerate transformers_stream_generator tiktoken einops
+```
+
+Execute the following python code:
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("pfnet/nekomata-7b-pfn-qfin", trust_remote_code=True)
+
+# Use GPU with bf16 (recommended for supported devices)
+# model = AutoModelForCausalLM.from_pretrained("pfnet/nekomata-7b-pfn-qfin", device_map="auto", trust_remote_code=True, bf16=True)
+
+# Use GPU with fp16
+# model = AutoModelForCausalLM.from_pretrained("pfnet/nekomata-7b-pfn-qfin", device_map="auto", trust_remote_code=True, fp16=True)
+
+# Use GPU with fp32
+# model = AutoModelForCausalLM.from_pretrained("pfnet/nekomata-7b-pfn-qfin", device_map="auto", trust_remote_code=True, fp32=True)
+
+# Use CPU
+# model = AutoModelForCausalLM.from_pretrained("pfnet/nekomata-7b-pfn-qfin", device_map="cpu", trust_remote_code=True)
+
+# Automatically select device and precision
+model = AutoModelForCausalLM.from_pretrained("pfnet/nekomata-7b-pfn-qfin", device_map="auto", trust_remote_code=True)
+
+text = "日本銀行は"
+input_ids = tokenizer(text, return_tensors="pt").input_ids.to(model.device)
+with torch.no_grad():
+  generated_tokens = model.generate(
+      inputs=input_ids,
+      max_new_tokens=32,
+      do_sample=True,
+      temperature=1.0,
+      repetition_penalty=1.1
+  )[0]
+generated_text = tokenizer.decode(generated_tokens)
+print(generated_text)
+# 日本銀行は、2016年9月に「長短金利操作付き量的・質的金融緩和」を導入し、長期国
+```
+
+## Model Details
+- Model size: 7b
+- Fine-tuned tokens: 370M tokens (Japanese: 300M tokens, English: 13M tokens, Digits: 14M tokens)
+- Context length: 2048
+- Developed by: Preferred Networks, Inc
+- Model type: Causal decoder-only
+- Language(s): Japanese and English
+- License: [Tongyi Qianwen LICENSE AGREEMENT](https://github.com/QwenLM/Qwen/blob/e8e15962d897714944773cca57fa2e460a3655e8/Tongyi%20Qianwen%20LICENSE%20AGREEMENT)
+
+## Bias, Risks, and Limitations
+nekomata-7b-pfn-qfin is a new technology that carries risks with use.
+Testing conducted to date has been in English and Japanese, and has not covered, nor could it cover all scenarios.
+For these reasons, as with all LLMs, nekomata-7b-pfn-qfin’s potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts.
+This model is not designed for legal, tax, investment, financial, or other advice.
+Therefore, before deploying any applications of nekomata-7b-pfn-qfin, developers should perform safety testing and tuning tailored to their specific applications of the model.
+
+## How to cite
+```
+@misc{hirano2024,
+      title={Construction of Domain-specified Japanese Large Language Model for Finance through Continual Pre-training}, 
+      author={Masanori Hirano and Kentaro Imajo},
+      year={2024},
+      eprint={2404.10555},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+```
+
+## Contributors
+Preferred Networks, Inc.
+ - Masanori Hirano
+ - Kentaro Imajo
+
+# License
+[Tongyi Qianwen LICENSE AGREEMENT](https://github.com/QwenLM/Qwen/blob/e8e15962d897714944773cca57fa2e460a3655e8/Tongyi%20Qianwen%20LICENSE%20AGREEMENT)