Skip to content

feat(train_reward_model): add chatml formatting and aggregation of more statistics #21

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
maxreciprocate wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat(train_reward_model): add chatml formatting and aggregation of more statistics #21

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8021879
feat(train_reward_model): force chatml & add stats
maxreciprocate Dec 7, 2023
9ae61bd
feat(README): usage snippet with `apply_chat_template`
maxreciprocate Dec 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 28 additions & 12 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,35 +2,51 @@
A repository for transformer critique learning and generation.

## Scalar reward models
Train [OpenLLaMA-13B](https://github.com/openlm-research/open_llama) on [Helpful and Harmless dataset](https://github.com/anthropics/hh-rlhf):
Train [Mistral-7B](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1) on [UltraFeedback](https://huggingface.co/datasets/allenai/ultrafeedback_binarized_cleaned) dataset:

```bash
accelerate launch --config_file configs/accelerate/zero2.yaml \
train_reward_model.py \
--model_path openlm-research/open_llama_13b \
--dataset pvduy/rm_oa_hh \
--batch_size 1 \
--model_path mistralai/Mistral-7B-Instruct-v0.1 \
--dataset allenai/ultrafeedback_binarized_cleaned:train_prefs \
--batch_size 4 \
--eval_interval 1000 \
--lr 0.00001 \
--lr 0.000003 \
--weight_decay 0 \
--num_unfrozen_layers 12 \
--gradient_checkpointing \
--checkpoint_dir checkpoints \
--calibration_datasets reciprocate/vicuna-fair-eval
--calibration_datasets allenai/ultrafeedback_binarized_cleaned:test_prefs Intel/orca_dpo_pairs reciprocate/fair-eval
```

Usage:

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from transformers import pipeline

reward_fn = pipeline(
"text-classification",
model="reciprocate/mistral-7b-rm",
truncation=True,
max_length=4096,
function_to_apply="none"
)

ckpt = "reciprocate/openllama-13b_rm_oasst-hh"
model = AutoModelForSequenceClassification.from_pretrained(ckpt, load_in_4bit=True)
tokenizer = AutoTokenizer.from_pretrained(ckpt)
chats = [[
{"role": "user", "content": "When was the battle at Waterloo?"},
{"role": "assistant", "content": "I think it was in 1983, but please double-check that when you have a chance."}
], [
{"role": "user", "content": "When was the battle at Waterloo?"},
{"role": "assistant", "content": "The battle at Waterloo took place on June 18, 1815."}
]]

model(**tokenizer("ASSISTANT: This sentence is a lie.", return_tensors="pt"))[0].item()
inputs = [reward_fn.tokenizer.apply_chat_template(chat, tokenize=False) for chat in chats]
output = reward_fn(inputs)
scores = [x["score"] for x in output]
scores
```

Output:
```python
-1.626953125
>>> [-1.0530743598937988, 0.6916144490242004]
```
Loading