Comparing two training datasets using sqlite-utils

WizardLM is "an Instruction-following LLM Using Evol-Instruct". It's a fine-tuned model on top of Meta's LLaMA. The fine-tuning uses 70,000 instruction-output pairs from this JSON file:

https://huggingface.co/datasets/victor123/evol_instruct_70k

Those instruction pairs were created using OpenAI's gpt-3.5-turbo API, using a process they call "prompt rewriting", which is described in full in their paper (PDF) (scroll to appendix A.1 for the prompts they used for this).

WizardLM-13B-Uncensored is a retrained version of WizardML described like so:

This is WizardLM trained with a subset of the dataset - responses that contained alignment / moralizing were removed. The intent is to train a WizardLM that doesn't have alignment built-in, so that alignment (of any sort) can be added separately with for example with a RLHF LoRA.

The training data for that is WizardLM_alpaca_evol_instruct_70k_unfiltered - another JSON file, this time with 54,974 rows.

I decided to take a look at the 70,000 - 54,974 = 15,026 rows that had been removed to see what those looked like.

Fetching the data

Both JSON files can be downloaded from Hugging Face:

wget https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered/resolve/main/WizardLM_alpaca_evol_instruct_70k_unfiltered.json
wget https://huggingface.co/datasets/victor123/evol_instruct_70k/resolve/main/alpaca_evol_instruct_70k.json

Resulting files:

95M    WizardLM_alpaca_evol_instruct_70k_unfiltered.json
130M   alpaca_evol_instruct_70k.json

The naming here is a bit confusing: the "unfiltered" file is the one with 54,974 rows (which has had the "moralizing" instructions filtered out of it).

I stuck with that confusing naming for the rest of this analysis.

Loading the data into SQLite

Both files are JSON arrays that look like this:

[
    {
        "instruction": "...",
        "output": "...",
    },
    {
        "instruction": "...",
        "output": "...",
    }
]

My sqlite-utils tool accepts JSON arrays like this and loads them into tables.

I ran this:

sqlite-utils insert everything.db unfiltered WizardLM_alpaca_evol_instruct_70k_unfiltered.json
sqlite-utils insert everything.db filtered alpaca_evol_instruct_70k.json

Now I can run sqlite-utils schema everything.db to get back this:

CREATE TABLE [unfiltered] (
   [instruction] TEXT,
   [output] TEXT
);
CREATE TABLE [filtered] (
   [instruction] TEXT,
   [output] TEXT
);

I want to see the rows in filtered that are not present in unfiltered.

I could probably have just done this with straight string comparison, but since the columns are quite long I decided to compare using MD5 hashes of the instruction columns instead.

I used sqlite-utils convert to add an md5 column containing the MD5 hash of the instruction to both tables:

sqlite-utils convert everything.db filtered instruction \
  'hashlib.md5(value.encode()).hexdigest()' \
  --import hashlib --output md5

sqlite-utils convert everything.db unfiltered instruction \
  'hashlib.md5(value.encode()).hexdigest()' \
  --import hashlib --output md5

The schema is now:

CREATE TABLE [unfiltered] (
   [instruction] TEXT,
   [output] TEXT,
   [md5] TEXT
);
CREATE TABLE [filtered] (
   [instruction] TEXT,
   [output] TEXT,
   [md5] TEXT
);

I opened this up in Datasette Desktop to check that everything looked good and run some exploratory queries.

In particular, this query showed me the rows that were missing from the unfiltered table:

select * from filtered where md5 not in (
    select md5 from unfiltered
)

Persisting that as a column

I decided to add a removed column to the filtered table, which would be 1 for rows that were missing from unfiltered and 0 for rows that were present in both.

First I added the blank column:

sqlite-utils add-column everything.db filtered removed integer

Then I set it to 0 for everything:

sqlite-utils everything.db 'update filtered set removed = 0'

Then I set it to 1 for the rows that were missing from unfiltered:

sqlite-utils everything.db '
  update filtered set removed = 1
  where md5 not in (select md5 from unfiltered)
'

That output the following:

[{"rows_affected": 15026}]

Exporting the data

I wanted to explore the data in Datasette Lite, so I exported it to JSON like this:

sqlite-utils everything.db \
  'select instruction, output from filtered where removed = 1' \
  > removed.json

The resulting file is 34MB and contains 15,026 rows.

Opening that in Datasette Lite

I added the new JSON file to this GitHub Gist to ensure it would have open CORS headers. Then I figured out the final raw URL to it:

https://gist.githubusercontent.com/simonw/8d6ecac3c55058b97e118e2e694a6234/raw/8fafc383f400e201c3eef1ac39a60e62eb9a2e46/removed.json

And opened it in Datasette Lite like this:

https://lite.datasette.io/?json=https://gist.githubusercontent.com/simonw/8d6ecac3c55058b97e118e2e694a6234/raw/8fafc383f400e201c3eef1ac39a60e62eb9a2e46/removed.json#/data/removed

What I found

In exploring the data using the following SQL query:

select * from removed order by random() limit 10

I quickly started to see examples of instructions that looked to me like they had been accidentally excluded. Things like this:

It looked like the filtering had been pretty basic - was it really just omitting any response with "sorry" in it?

It turns out it was! I found the script that was used for the filtering here:

https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered/blob/main/wizardlm_clean.py

It's a big list of strings to exclude, which includes:

"I cannot assist",
"prioritize ethical",
"respectful",
"morally",
"I'm sorry,",
"I'm an",
"I am an",
"I'm an AI" ,
"I am an AI",
"my purpose",

If I'd found that Python file a little bit earlier it would have saved me some enjoyable mucking around with SQLite and JSON!

Related

• duckdb Use DuckDB to convert parquet to JSON and then open it in Datasette Lite - 2023-03-21
• llms Storing and serving related documents with openai-to-sqlite and embeddings - 2023-08-14
• python Calculating embeddings with gtr-t5-large in Python - 2023-01-31
• datasette Interactive row selection prototype with Datasette - 2023-03-30
• llms Embedding paragraphs from my blog with E5-large-v2 - 2023-09-08
• sqlite Sort by number of JSON intersections - 2022-08-17
• bash Escaping a SQL query to use with curl and Datasette - 2020-12-08
• googlecloud Analyzing Google Cloud spend with Datasette - 2022-08-16
• sqlite Combining substr and instr to extract text - 2022-02-15
• llms Summarizing Hacker News discussion themes with Claude and LLM - 2023-09-09

Created 2023-05-23T16:01:22-07:00 · Edit