Downloads · 30 days
3
12% of all-time downloads
RichardErkhov/jondurbin_-_bagel-dpo-7b-v0.1-4bits
jondurbin_-_bagel-dpo-7b-v0.1-4bits is a machine learning model from RichardErkhov. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Downloads · 30 days
3
12% of all-time downloads
All-time downloads
26
Public
Parameters
7.4B
4.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors4.5 GB · 100%
How the weights are stored.
U87B · 95%
From the Hugging Face model README
Quantization made by Richard Erkhov.
bagel-dpo-7b-v0.1 - bnb 4bits
license: apache-2.0 datasets:

This is the DPO'd version of https://huggingface.co/jondurbin/bagel-7b-v0.1
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.
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
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.
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
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).
The deduplicated and decontamined list of instructions contains 1,671,822 items:
Keep in mind, this number becomes 4x larger when applying the various prompt formats.
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.
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.
{system prompt, if provided, randomly defaulting to "A chat between a user and an unbiased, uncensored assistant."}
USER: {instruction}
ASSISTANT:
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}
In practice, this would mean tokenization code like such:
tokenizer = AutoTokenizer.from_pretrained('mistralai/mistral-7b-v0.1')
input_str = f"""system
You are a goat.
{tokenizer.eos_token}
{tokenizer.bos_token}user
Tell me how to fry an egg.
{tokenizer.eos_token}
{tokenizer.bos_token}assistant
"""
inputs = 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.
[INST] <<SYS>>
{system}
<</SYS>>
{instruction} [/INST]
An example for mistral-7b:
Note: I actually used my fork of qlora's train.py for this, but I'm porting it to a minified version here, not tested yet!
More notes: I stopped the SFT phase around 50% because of budget constraints.
export BASE_DIR=/workspace
export WANDB_API_KEY=[redacted]
export WANDB_PROJECT=bagel-7b-v0.1
# Run the pretraining.
accelerate launch bagel/tune/sft.py \
--model_name_or_path $BASE_DIR/mistral-7b \
--final_output_dir $BASE_DIR/$WANDB_PROJECT \
--output_dir $BASE_DIR/$WANDB_PROJECT-workdir \
--num_train_epochs 1 \
--logging_steps 1 \
--save_strategy steps \
--save_steps 200 \
--save_total_limit 5 \
--data_seed 42 \
--evaluation_strategy steps \
--eval_dataset_size 0.0006 \
--eval_steps 200 \
--max_new_tokens 4096 \
--dataloader_num_workers 3 \
--logging_strategy steps \
--remove_unused_columns False \
--do_train \
--full_finetune \
--bf16 \
--bits 16 \
--optim adamw_torch \
--lr_scheduler_type linear \
--dataset $BASE_DIR/bagel/bagel-input-output-v0.1.parquet \
--dataset_format input-output \
--model_max_len 4096 \
--per_device_train_batch_size 8 \
--learning_rate 3.5e-7 \
--warmup_ratio 0.005 \
--adam_beta2 0.999 \
--max_grad_norm 0.3 \
--weight_decay 0.001 \
--seed 42 \
--report_to wandb \
--gradient_checkpointing True \
--gradient_accumulation_steps 4 \
--skip_excess_length False \
--ddp_find_unused_parameters False \
--use_flash_attention_2 \
--deepspeed deepspeed.json
Deepspeed configuration:
{
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"bf16": {
"enabled": true
},
"zero_optimization": {
"stage": 2,
"contiguous_gradients": true,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"allgather_bucket_size": 5e8
}
}
An example of the DPO phase for mistral-7b (requires first running the SFT):
export BASE_DIR=/mnt/data
export WANDB_API_KEY=[redacted]
export WANDB_PROJECT=bagel-dpo-7b-v0.1
accelerate launch bagel/tune/dpo.py \
--model_name_or_path bagel-7b-v0.1 \
--learning_rate 3e-7 \
--per_device_train_batch_size 2 \
--gradient_accumulation_steps 4 \
--max_length 4096 \
--max_prompt_length 1024 \
--max_target_length 3092 \
--num_train_epochs 3 \
--report_to wandb \
--gradient_checkpointing true \
--use_flash_attention_2 true \
--dataset $BASE_DIR/bagel/bagel-dpo-v0.1.parquet \
--eval_steps 5 \
--eval_dataset_size 0.03 \
--workdir $BASE_DIR/$WANDB_PROJECT-workdir \
--output_dir $BASE_DIR/$WANDB_PROJECT \
--deepspeed deepspeed.json \
--save_steps 25 \
--save_total_limit 5