File size: 1,920 Bytes
17fb6e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
---
language:
- en
license: mit
task_categories:
- question-answering
- table-question-answering
tags:
- finance
- conversational-qa
- numerical-reasoning
- financial-tables
- 10-k
pretty_name: ConvFinQA (Tomoro pre-cleaned)
size_categories:
- 1K<n<10K
---

# ConvFinQA (Tomoro pre-cleaned)

Re-host of the pre-cleaned ConvFinQA dataset distributed as part of Tomoro AI's "Applied AI Solution Engineer" take-home exercise. ConvFinQA itself is the conversational QA benchmark over single-page financial documents from Chen et al. (EMNLP 2022).

The data is a single JSON file with two splits, `train` (3,037 records) and `dev` (421 records). Each record carries:

- `id`: stable per-record identifier
- `doc`: a 10-K page split into `pre_text`, `post_text`, and a column-keyed `table` (nested dict: column header → row header → cell value)
- `dialogue`: `conv_questions`, `conv_answers`, `turn_program` (FinQA DSL), `executed_answers` (the dataset's gold executed values, numeric or `'yes'`/`'no'`), and `qa_split` (origin flag for hybrid two-source dialogues)
- `features`: precomputed flags (`num_dialogue_turns`, `has_type2_question`, `has_duplicate_columns`, `has_non_numeric_values`)

## Loading

```python
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="sharick008/convfinqa",
    filename="convfinqa_dataset.json",
    repo_type="dataset",
)
data = json.load(open(path))
print(len(data["train"]), len(data["dev"]))
```

## Citation

> Chen, Z., Li, S., Smiley, C., Ma, Z., Shah, S., & Wang, W. Y. (2022). ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering. EMNLP.

## Licence

ConvFinQA upstream is MIT-licensed. This re-host preserves that licence and the original attribution to the ConvFinQA authors. Tomoro's pre-cleaning passes (column disambiguation, scale normalisation, numeric coercion) are applied on top.