File size: 4,308 Bytes
a6feb32
 
 
c25607a
 
 
da87cde
c25607a
da87cde
b4602ef
c25607a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
53
54
55
56
57
---
license: cc-by-sa-3.0
---

### Summary

`databricks-dolly-15k-cleanset` can be used to produced CLEANed up versions of the popular [databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) dataSET, which was used to fine-tune the [Dolly 2.0](https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm). The original `databricks-dolly-15k` contains 15,000 human-annotated instruction-response pairs covering various categories. However, there are many low-quality responses, incomplete/vague prompts, and other problematic text lurking in the dataset (as with for all real-world instruction tuning datasets). We ran Cleanlab Studio to automatically detect low quality datapoints in the original dataset. Our `databricks-dolly-15k-cleanset` appends the following columns to the original dataset, which are various data quality measures from Cleanlab:

- `TLM_confidence_score`: A measure of the trustworthiness of a response to a given prompt (counts for both *aleatoric and epistemic uncertainties).* Represented by a value between 0 and 1, with lower values indicating the response is unlikely to be good.
- `cleanlab_PII_score`: A measure of the occurrence and severity of Personally Identifiable Information (PII) within the text. Represented by a value between 0 and 1, with higher values indicating greater severity.
- `cleanlab_informal_score`: A measure of the occurrence and severity of casual language, slang, or poor writing within the text. Represented by a value between 0 and 1, with higher values indicating greater severity.
- `cleanlab_non_english_score`: A measure of the occurrence of text written in a foreign language or containing nonsensical characters (such as HTML/XML tags, identifiers, hashes, random characters). Represented by a value between 0 and 1, with higher values indicating greater severity.
- `cleanlab_toxic_score`: A measure of the occurrence and severity of hateful speech and harmful language within the text. Represented by a value between 0 and 1, with higher values indicating greater severity.

Only a few lines of Cleanlab code are required to reproduce the `databricks-dolly-15k-cleanset` from the original `databricks-dolly-15k`, the code is available [here](https://github.com/cleanlab/cleanlab-tools/blob/main/fine_tuning_data_curation/fine_tuning_data_curation.ipynb).  

If you’re interested in learning how to detect bad data in your instruction tuning dataset for better LLM fine-tuning, check our our [blog](https://cleanlab.ai/blog/filter-llm-tuning-data/).

## Intended Uses

With the new columns, you can filter out low-quality datapoints to produce a cleaner dataset. If you have the time and resources, your can manually review the datapoints with problematic scores and replace them with higher quality instruction / responses. If not, you can determine thresholds for confidence and text issue scores, and automatically drop any datapoint whose scores falls on the wrong end of the thresholds, as shown below.

```python
import pandas as pd

# Load the dataset
df = pd.read_csv('databricks-dolly-15k-cleanset.csv')

# Lower confidence scores are more problematic
TLM_confidence_score_threshold = 0.5

# Higher text issues scores are more problematic
PII_score_threshold = 0.4
informal_score_threshold = 0.6
non_english_score_threshold = 0.8
toxic_score_threshold = 0.95

cleaned_df = df[
    (df['TLM_confidence_score'] > TLM_confidence_score_threshold) &
    (df['cleanlab_PII_score'] < PII_score_threshold) &
    (df['cleanlab_informal_score'] < informal_score_threshold) &
    (df['cleanlab_non_english_score'] < non_english_score_threshold) &
    (df['cleanlab_toxic_score'] < toxic_score_threshold)
]

# Drop the score columns
columns_to_drop = ['TLM_confidence_score', 'cleanlab_PII_score', 'cleanlab_informal_score', 
                   'cleanlab_toxic_score', 'cleanlab_non_english_score']
cleaned_df = cleaned_df.drop(columns=columns_to_drop)

# Save to file. We now have a clean version of the original dataset.
cleaned_df.to_csv('databricks-dolly-15k-cleaned.csv', index=False)

```

We have provided one such cleaned version of the dataset here: 
[databricks-dolly-15k-cleaned.csv](https://huggingface.co/datasets/Cleanlab/databricks-dolly-15k-cleaned)