license: other
license_name: qwen
license_link: https://huggingface.co/Qwen/Qwen-14B/blob/main/LICENSE
datasets:
- ai2_arc
- unalignment/spicy-3.1
- codeparrot/apps
- facebook/belebele
- boolq
- jondurbin/cinematika-v0.1
- drop
- lmsys/lmsys-chat-1m
- TIGER-Lab/MathInstruct
- cais/mmlu
- Muennighoff/natural-instructions
- openbookqa
- piqa
- Vezora/Tested-22k-Python-Alpaca
- cakiki/rosetta-code
- Open-Orca/SlimOrca
- spider
- squad_v2
- migtissera/Synthia-v1.3
- datasets/winogrande
- nvidia/HelpSteer
- Intel/orca_dpo_pairs
- unalignment/toxic-dpo-v0.1
- jondurbin/truthy-dpo-v0.1
- allenai/ultrafeedback_binarized_cleaned
- Squish42/bluemoon-fandom-1-1-rp-cleaned
- LDJnr/Capybara
- JULIELab/EmoBank
- kingbri/PIPPA-shareGPT
A bagel, with everything
Overview
An experimental fine-tune of qwen-14b using bagel
The resulting model didn't turn out quite as great as I would have liked - in fact, I'd probably use the mistral-7b version over this, because it scored higher on mt-bench, is much faster, and generally is uncensored in comparison to this model (even after toxic DPO, several epochs)
I modified the qwen tokenizer to use <s> instead of <|im_start|> and </s> instead of <|endoftext|>, and it may have caused some issues but I'm not entirely sure.
Hardware kindly provided by Massed Compute
Data selection.
The first step in the process is creating a dataset. In this case, we're actually creating a composite dataset, consisting of both supervised fine-tuning data (SFT) and direct preference optimization (DPO) data.
All instruction data, that is, data that is not plain text (like project Gutenberg and items from Cinematika) or DPO, is converted into ShareGPT format so it's easier to work with.
See the corresponding code in bagel/data_sources/*.py for full implementation for each data source.
Deduplication is done by creating a uuid v5 of the instruction/text, then only adding items not previously seen (where datasets are loaded in order of the confidence score I assign them). This means that if an instruction is in data source "Foo" with confidence 4 as well as in data source "Bar" with confidence score 2, only the entry from "Foo" will be taken.
SFT data sources
Yes, you will see benchmark names in the list, but this only uses the train splits, and a decontamination by cosine similarity is performed at the end as a sanity check
- ai2_arc
- Abstraction and reasoning dataset, useful in measuring "intelligence" to a certain extent.
- airoboros
- Variety of categories of synthetic instructions generated by gpt-4.
- apps
- Python coding dataset with 10k problems.
- belebele
- Multi-lingual reading comprehension dataset.
- bluemoon
- Roleplay data scraped from Bluemoon, then cleaned and formatted as ShareGPT.
- boolq
- Corpus of yes/no questions (which can be surprisingly difficult for AI to answer apparently?)
- capybara
- Multi-turn dataset used to create the capybara models.
- cinematika (instruction and plain text)
- RP-style data synthesized from movie scripts so the model isn't quite as boring as it otherwise would be.
- drop
- More reading comprehension.
- emobank
- Emotion annotations using the Valence-Arousal-Domninance scheme.
- gutenberg (plain text)
- Books/plain text, again to make the model less boring, only a handful of examples supported by chapterize
- lmsys_chat_1m (only gpt-4 items, also used for DPO)
- Chats collected by the lmsys chat arena, containing a wide variety of chats with various models.
- mathinstruct
- Composite dataset with a variety of math-related tasks and problem/question formats.
- mmlu
- Massive Multitask Language Understanding - a wide variety of questions about various subject matters.
- natural_instructions
- Millions of instructions from 1600+ task categories (sampled down substantially, stratified by task type)
- openbookqa
- Question answering dataset.
- pippa
- Deduped version of PIPPA in ShareGPT format.
- piqa
- Phyiscal interaction question answering.
- python_alpaca
- Python instruction response pairs, validated as functional.
- rosetta_code
- Code problems and solutions in a variety of programming languages taken from rosettacode.org.
- slimorca
- Collection of ~500k gpt-4 verified chats from OpenOrca.
- spider
- SQL-targeted dataset.
- squad_v2
- Contextual question answering (RAG).
- synthia
- GPT-4 generated data using advanced prompting from Migel Tissera.
- winogrande
- Fill in the blank style prompts.
DPO data sources
- airoboros 3.1 vs airoboros 2.2.1
- The creative/writing tasks from airoboros-2.2.1 were re-generated using gpt4-0314 and a custom prompt to get longer, more creative, less clichè responses for airoboros 3.1, so we can use the shorter/boring version as the "rejected" value and the rerolled response as "chosen"
- helpsteer
- Really neat dataset provided by the folks at NVidia with human annotation across a variety of metrics. Only items with the highest "correctness" value were used for DPO here, with the highest scoring output as "chosen" and random lower scoring value as "rejected"
- orca_dpo_pairs
- Another interesting dataset by Intel, which provides various DPO pairs generated from prompts included in the SlimOrca dataset.
- toxic-dpo
- highly toxic and potentially illegal content! De-censorship, for academic and lawful purposes only, of course. Generated by llama-2-70b via prompt engineering.
- truthy
- DPO pairs meant to increase truthfulness of the model, e.g. common misconceptions, differentiate between AI assistants and roleplayed human in terms of corporeal awareness/locality/etc.
- ultrafeedback
- One of the bits of magic behind the Zephyr model. Only the items with a chosen score of 8 or higher were included.
Only the train splits were used (if a split was provided), and an additional pass of decontamination is performed using approximate nearest neighbor search (via faiss).
Prompt formatting
In sticking with the theme of the bagel, I didn't want to use a single prompt format, so I used 4 - vicuna, llama-2, alpaca, and chat-ml (sorta). I also didn't want to randomly select a single prompt format for each item (hoping each instruction would generalize more when used in a variety of prompt formats), so each instruction is actually converted into every prompt format.
This means each epoch of our fine-tune is really basically 4 epochs. So, for the fine-tunes, I would recommend only doing 1 epoch (or 0.75 epochs). I am testing with a single epoch using a relatively low learning rate.
Alpaca (sort of)
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{system prompt, if provided}
{instruction}
### Response:
The main difference here is that because of the dataset formatting and variety of data sources, it would have been much to tedious to add an ### Input: block, so the inputs are just in the instruction section.
Vicuna
{system prompt, if provided, randomly defaulting to "A chat between a user and an unbiased, uncensored assistant."}
USER: {instruction}
ASSISTANT:
ChatML (sort of)
I don't really understand the point of having special tokens for <|im_start|> and <|im_end|>, because in practice they just act as BOS and EOS tokens (but, please correct me if I'm wrong).
So, instead of:
{bos}<|im_start|>{role}
{text}
<|im_end|>{eos}
I just changed it to:
{bos}{role}
{text}
{eos}
If you really want to use <|im_start|> and <|im_end|>, just update your tokenizer_config.json to use <|im_start|> instead of <s> and <|im_end|> instead of </s> and when tokenizing. And if you still don't like what I've done to this chat-ml-ish format, feel free to cry into your pillow or fork the code and do a new fine-tune.
Llama-2 chat
[INST] <<SYS>>
{system}
<</SYS>>
{instruction} [/INST]