If you are getting too many AALLM or other refusals, even with explicitly human system prompts, you may want to try the non-DPO version.
Benchmarks
I ran these against the latest main branch of lm-evaluation-harness (and opencompass/FastChat for agieval and mt-bench), since batch size/etc effects score for some benchmarks.
model
arc_challenge
boolq
gsm8k
hellaswag
mmlu
openbookqa
piqa
truthful_qa
winogrande
bagel
0.6715
0.8813
0.5618
0.8397
0.6408
0.51
0.8406
0.6275
0.7561
openhermes-2.5
0.6476
0.8835
0.4852
0.8414
0.6347
0.498
0.8400
0.5295
0.7443
MT-Bench:
########## First turn ##########
score
model turn
bagel-7b-v0.1 1 7.60625
########## Second turn ##########
score
model turn
bagel-7b-v0.1 2 7.00625
########## Average ##########
score
model
bagel-7b-v0.1 7.30625
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
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"
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"
highly toxic and potentially illegal content! De-censorship, for academic and lawful purposes only, of course. Generated by llama-2-70b via prompt engineering.
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.
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).
Total dataset size
The deduplicated and decontamined list of instructions contains 1,671,822 items:
1,602,217 SFT/instructions
59,247 DPO pairs
1606 with both SFT and DPO data
Keep in mind, this number becomes 4x larger when applying the various prompt formats.
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:
text
1{bos}<|im_start|>{role}
2{text}
3<|im_end|>{eos}
I just changed it to:
text
1{bos}{role}
2{text}
3{eos}
In practice, this would mean tokenization code like such:
python
1tokenizer = AutoTokenizer.from_pretrained('mistralai/mistral-7b-v0.1')23input_str =f"""system
4You are a goat.
5{tokenizer.eos_token}6{tokenizer.bos_token}user
7Tell me how to fry an egg.
8{tokenizer.eos_token}9{tokenizer.bos_token}assistant
10"""1112inputs = tokenizer(input_str, return_tensors="pt")
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.
