---
language: 
  - en
license: mit
tags:
  - finance
  - custom-model
  - pytorch
  - conversational-qa
  - financial-qa
datasets:
  - conv_finqa
---

# TinyRecursiveModel for ConvFinQA

## Model Details

### Model Description

This model is a custom **TinyRecursiveModel (TRM)** fine-tuned specifically for the **ConvFinQA** dataset. It is designed to handle conversational question answering over complex financial documents, earnings reports, and tables. 

* **Developed by:** KenyaWashed
* **Model type:** Custom PyTorch Model (`TinyRecursiveModel`)
* **Language(s) (NLP):** English
* **License:** MIT
* **Finetuned from model / Base Tokenizer:** [Điền tên base tokenizer, ví dụ: `roberta-base` hoặc `ProsusAI/finbert`]

## Uses

### Direct Use

The model is built for researchers and developers working in Financial NLP. It can be used to extract answers and perform hierarchical reasoning over financial texts and tables in a conversational context.

### Out-of-Scope Use

This model is not intended to provide professional financial advice or real-time trading signals. It is a research artifact focused on natural language processing and reasoning.

## How to Get Started with the Model

Since this model uses a custom `TinyRecursiveModel` architecture, you will need the original class definition in your codebase to load the weights properly.

Here is how you can load the tokenizer and the model weights:

```python
import torch
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download
# Nhớ import class TinyRecursiveModel từ source code của mày
# from your_custom_module import TinyRecursiveModel 

repo_id = "KenyaWashed/trm-convfinqa"

# 1. Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(repo_id)

# 2. Khởi tạo model base (nhớ truyền đúng tham số lúc train)
model = TinyRecursiveModel(
    # [Điền các tham số khởi tạo model của mày vào đây]
)

# 3. Download weights từ Hugging Face và load vào model
model_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin")
model.load_state_dict(torch.load(model_path, map_location="cpu"))
model.eval()

print("Model loaded successfully!")