metadata

license: apache-2.0
base_model: TinyLlama/TinyLlama_v1.1
tags:
  - trl
  - reward-trainer
  - generated_from_trainer
metrics:
  - accuracy
model-index:
  - name: tinyllama_rm_sentiment_1b
    results: []
datasets:
  - trl-internal-testing/sentiment-trl-style

tinyllama_rm_sentiment_1b

This model is a fine-tuned version of TinyLlama/TinyLlama_v1.1 on https://huggingface.co/datasets/trl-internal-testing/sentiment-trl-style. It achieves the following results on the evaluation set:

Loss: 0.6514
Accuracy: 0.625

Model description

Trained using:

python trl/examples/scripts/rm/rm.py \
--dataset_name trl-internal-testing/sentiment-trl-style \
--dataset_train_split train \
--dataset_eval_split test \
--model_name_or_path TinyLlama/TinyLlama_v1.1 \
--chat_template simple_concat \
--learning_rate 3e-6 \
--per_device_train_batch_size 32 \
--per_device_eval_batch_size 32 \
--gradient_accumulation_steps 1 \
--logging_steps 1 \
--eval_strategy steps \
--max_token_length 1024 \
--max_prompt_token_lenth 1024 \
--remove_unused_columns False \
--num_train_epochs 1 \
--eval_steps 100 \
 --output_dir models/ppo_torchtune/tinyllama/tinyllama_rm_sentiment_1b \
 --push_to_hub

on the "dataset-processor" branch of trl:

git clone -b "dataset-processor" https://github.com/huggingface/trl

Intended uses & limitations

More information needed

Training and evaluation data

https://huggingface.co/datasets/trl-internal-testing/sentiment-trl-style

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-06
train_batch_size: 32
eval_batch_size: 32
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.6033	0.6410	100	0.6514	0.625

Framework versions

Transformers 4.42.2
Pytorch 2.2.0+cu121
Datasets 2.20.0
Tokenizers 0.19.1