Model Card for plutus-8B-instruct

This model is an instruction-tuned large language model specialized for Greek financial texts. It is based on the ilsp/Llama-Krikri-8B-Instruct model and has been fine-tuned using Low-Rank Adaptation (LoRA) on a mixed Greek financial dataset.

Model Details

Model Description

This model card describes plutus-8B-instruct, a model developed by The Fin AI and finetuned to serve Greek-centric financial language tasks. The model leverages parameter-efficient fine-tuning (PEFT) via LoRA and is designed to generate or understand financial texts in Greek. The training was carried out with a local command-line backend with logging managed via TensorBoard.

• Developed and Funded by: NaCTeM, Archimedes RC, The Fin AI
• Model type: Instruction-tuned large language model, specialized for low-resource domains such as Greek finance
• Language(s) (NLP): Greek
• License: LlaMA 3.1 Community License
• Finetuned from model [optional]: ilsp/Llama-Krikri-8B-Instruct

Model Sources

• Repository: https://huggingface.co./TheFinAI/plutus-8B-instruct
• Paper: https://arxiv.org/abs/2502.18772
• Demo: https://huggingface.co./spaces/TheFinAI/plutus-8B-instruct

Uses

Direct Use

Plutus-8B-instruct can be directly applied in Greek finance applications such as answering user queries, summarizing financial reports, or generating context-aware financial planning text. Users should ensure that input texts align with the financial domain to achieve optimal performance.

Downstream Use

The model can be integrated into larger systems such as chatbots, recommendation systems, or data analysis pipelines focusing on Greek financial markets. Fine-tuning on domain-specific datasets might further improve performance for specialized tasks.

Out-of-Scope Use

This model is not designed for non-financial applications or for languages other than Greek, and may not perform reliably if used outside of the Greek financial context. It should not be used for high-stakes financial decision-making without additional verification.

Bias, Risks, and Limitations

As with many LLMs, plutus-8B-instruct may exhibit biases present in the training data, and its outputs may require human review in sensitive contexts. Users should be aware that the model may generate plausible-sounding yet factually inaccurate or biased financial advice.

Recommendations

Users (both direct and downstream) should critically evaluate the outputs in the context of local regulations and financial best practices. It is recommended to apply additional validation measures when deploying the model in production environments.

How to Get Started with the Model

To load and run plutus-8B-instruct, you can use the Hugging Face Transformers library. For example:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "TheFinAI/plutus-8B-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Παρακαλώ δώσε μου ανάλυση για την οικονομική κατάσταση της Ελλάδας."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Additional details for running with mixed precision, LoRA configuration, or int4 quantization can be found in the training documentation.

Training Details

Training Data

The model was fine-tuned using the dataset available at https://huggingface.co./collections/TheFinAI/plutus-benchmarking-greek-financial-llms-67bc718fb8d897c65f1e87db. The training split ("train") was used, and the dataset contains various Greek financial texts.

Training Procedure

The finetuning process used parameter-efficient fine-tuning (PEFT) with LoRA. The base model was ilsp/Llama-Krikri-8B-Instruct, and the training was executed on a local CLI backend with logs monitored via TensorBoard.

Preprocessing

Inputs were tokenized with a specialized tokenizer configured for the Greek language, and a chat template was applied to structure the conversational data accordingly. Padding was applied to the right as per the configuration.

Training Hyperparameters

• Block size: 4096
• Model max length: 42000
• Epochs: 3
• Batch size: 1
• Learning rate: 0.0005
• PEFT (LoRA): Enabled with:
  • lora_r: 16
  • lora_alpha: 32
  • lora_dropout: 0
• Quantization: int4
• Target Modules: all-linear
• Optimizer: adamw_torch
• Scheduler: cosine
• Gradient accumulation: 4
• Mixed precision: bf16

Speeds, Sizes, Times

Training was performed on locally available hardware. Specific details on throughput, training runtime, and checkpoint sizes will be provided in subsequent documentation.

Evaluation

Testing Data, Factors & Metrics

Testing Data

Evaluation was performed on reserved portions of the training dataset and additional financial texts. For detailed information, users should refer to the corresponding Dataset Card for https://huggingface.co./collections/TheFinAI/plutus-benchmarking-greek-financial-llms-67bc718fb8d897c65f1e87db

Factors

The evaluation considered multiple aspects including:

• Domain-specific performance (Greek financial texts)
• General language understanding within Greek
• Response coherence and factual relevance

Metrics

Evaluation metrics included standard language generation measures as well as domain-specific qualitative assessments. Further discussion on the metrics used will be provided in future updates.

Results

Evaluated across multiple Greek financial NLP tasks. Bold values denote the highest scores, while underlined values indicate the second-highest scores in each column.

Model	GRFinNUM (Entity F1)	GRFinNER (Entity F1)	GRFinQA (Acc)	GRFNS-2023 (Rouge-1)	GRMultiFin (Acc)	Mean
Open-source Small Models
LLaMA-3.2-1B	0.00	0.00	0.29	0.14	0.39	0.16
LLaMA-3-8b	0.00	0.13	0.33	0.07	0.70	0.25
LLaMA-3.1-8b	0.10	0.21	0.40	0.20	0.54	0.29
Qwen2.5-1.5B	0.00	0.00	0.36	0.02	0.31	0.14
Qwen2.5-7B	0.00	0.13	0.43	0.07	0.54	0.23
Gemma-2-2B	0.00	0.16	0.22	0.03	0.41	0.16
Gemma-2-9B	0.02	0.05	0.31	0.06	0.61	0.21
Mistral-7B	0.00	0.00	0.30	0.14	0.39	0.17
Open-source Large Models
Deepseek-V3	0.07	0.00	0.50	0.38	0.61	0.31
LLaMA-3-70B	0.05	0.45	0.60	0.08	0.61	0.36
Qwen2.5-32B	0.37	0.55	0.60	0.10	0.70	0.47
Qwen2.5-72B	0.32	0.39	0.74	0.04	0.72	0.44
Gemma-2-27B	0.18	0.18	0.25	0.09	0.61	0.26
Proprietary Models
GPT-3.5-Turbo	0.14	0.30	0.51	0.31	0.50	0.35
GPT-4o-Mini	0.25	0.30	0.12	0.36	0.59	0.32
GPT-4o	0.09	0.31	0.78	0.26	0.59	0.41
GPT-4	0.28	0.60	0.71	0.38	0.63	0.52
English Financial Models
Finma-7B	0.00	0.00	0.25	0.11	0.35	0.14
FinLLaMA-8B	0.00	0.00	0.28	0.03	0.38	0.14
Greek General Models
Meltemi-7B	0.12	0.50	0.48	0.19	0.43	0.34
Llama-Kríkri-8B	0.19	0.45	0.57	0.22	0.39	0.36
Greek Financial Models
Plutus-8B	0.70	0.57	0.64	0.34	0.72	0.60

Summary

In summary, plutus-8B-instruct is a domain-specific, instruction-tuned large language model optimized for Greek finance applications. Although it performs well in targeted scenarios, users are advised to validate its outputs carefully.

Citation

BibTeX:

@misc{peng2025plutusbenchmarkinglargelanguage,
      title={Plutus: Benchmarking Large Language Models in Low-Resource Greek Finance}, 
      author={Xueqing Peng and Triantafillos Papadopoulos and Efstathia Soufleri and Polydoros Giannouris and Ruoyu Xiang and Yan Wang and Lingfei Qian and Jimin Huang and Qianqian Xie and Sophia Ananiadou},
      year={2025},
      eprint={2502.18772},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.18772}, 
}

APA:

Peng, X., Papadopoulos, T., Soufleri, E., Giannouris, P., Xiang, R., Wang, Y., Qian, L., Huang, J., Xie, Q., & Ananiadou, S. (2025). Plutus: Benchmarking Large Language Models in Low-Resource Greek Finance. arXiv preprint arXiv:2502.18772. https://arxiv.org/abs/2502.18772

Glossary

• PEFT: Parameter-Efficient Fine-Tuning.
• LoRA: Low-Rank Adaptation, a technique to reduce the number of trainable parameters.
• BF16: bfloat16, a mixed precision format used to optimize training speed.
• Int4 Quantization: A lower precision format aimed at reducing model size and inference latency.

More Information

For more details regarding training logs, dataset preparations, and further technical insights, please refer to the associated GitHub repositories and documentation provided by The Fin AI.

Model Card Authors

The model card was prepared by The Fin AI with inputs from the Hugging Face community.

Model Card Contact

For additional questions or feedback, please contact The Fin AI.