license: llama3.1 language:

• en pipeline_tag: text-generation datasets:
• allenai/RLVR-GSM-MATH-IF-Mixed-Constraints base_model:
• allenai/Llama-3.1-Tulu-3-8B-DPO library_name: transformers

Llama-3.1-Tulu-3.1-8B

Tülu 3 is a leading instruction following model family, offering a post-training package with fully open-source data, code, and recipes designed to serve as a comprehensive guide for modern techniques. This is one step of a bigger process to training fully open-source models, like our OLMo models. Tülu 3 is designed for state-of-the-art performance on a diversity of tasks in addition to chat, such as MATH, GSM8K, and IFEval.

Version 3.1 update: The new version of our Tülu model is from an improvement only in the final RL stage of training. We switched from PPO to GRPO (no reward model) and did further hyperparameter tuning to achieve substantial performance improvements across the board over the original Tülu 3 8B model, as shown in the comparison below:

Model description

• Model type: A model trained on a mix of publicly available, synthetic and human-created datasets.
• Language(s) (NLP): Primarily English
• License: Llama 3.1 Community License Agreement
• Finetuned from model: allenai/Llama-3.1-Tulu-3-8B-DPO

Model Sources

• Training Repository: https://github.com/allenai/open-instruct
• Eval Repository: https://github.com/allenai/olmes
• Paper: https://arxiv.org/abs/2411.15124
• Demo: https://playground.allenai.org/

Model Family

Using the model

Loading with HuggingFace

To load the model with HuggingFace, use the following snippet:

from transformers import AutoModelForCausalLM

tulu_model = AutoModelForCausalLM.from_pretrained("allenai/Llama-3.1-Tulu-3.1-8B")

VLLM

As a Llama base model, the model can be easily served with:

vllm serve allenai/Llama-3.1-Tulu-3.1-8B

Note that given the long chat template of Llama, you may want to use --max_model_len=8192.

Chat template

The chat template for our models is formatted as:

<|user|>\nHow are you doing?\n<|assistant|>\nI'm just a computer program, so I don't have feelings, but I'm functioning as expected. How can I assist you today?<|endoftext|>

Or with new lines expanded:

<|user|>
How are you doing?
<|assistant|>
I'm just a computer program, so I don't have feelings, but I'm functioning as expected. How can I assist you today?<|endoftext|>

It is embedded within the tokenizer as well, for tokenizer.apply_chat_template.

System prompt

In Ai2 demos, we use this system prompt by default:

You are Tulu 3, a helpful and harmless AI Assistant built by the Allen Institute for AI.

The model has not been trained with a specific system prompt in mind.

Bias, Risks, and Limitations

The Tülu3 models have limited safety training, but are not deployed automatically with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so). It is also unknown what the size and composition of the corpus was used to train the base Llama 3.1 models, however it is likely to have included a mix of Web data and technical sources like books and code. See the Falcon 180B model card for an example of this.

Performance

Note, see the updated version of the paper for the latest, fixed evaluations that improve scores for models such as Qwen 2.5 Instruct.

Hyperparamters

GRPO settings for RLVR:

• Learning Rate: 5 × 10⁻⁷
• Discount Factor (gamma): 1.0
• Mini-batches (N_mb): 2
• PPO-style Update Iteration (K): 1
• Clipping Coefficient (epsilon): 0.2
• Gradient Norm Threshold: 1.0
• Learning Rate Schedule: Constant
• Generation Temperature: 1.0
• Number of Samples per Prompt: 16
• Number of Unique Prompts per Training Iteration: 48
• Batch Size (effective): 48 * 16 = 768
• Max Token Length: 2,048
• Max Prompt Token Length: 2,048
• Penalty Reward Value for Responses without an EOS Token: 0.0
• Response Length: 2,048
• Total Episodes: 10,000,000 (the actual checkpoint is at episode 1474560)
• KL penalty coefficient (beta): 0.01
• Warm up ratio (omega): 0.0

Learning curves

Below is the training curves for Llama-3.1-Tulu-3.1-8B:

Below are the core eval scores over steps for Llama-3.1-Tulu-3.1-8B (note we took step 1920 as the final checkpoint, corresponding to episode 1,474,560):

Below are the other eval scores over steps for Llama-3.1-Tulu-3.1-8B (the codex evals had a bug and earlier scores are not shown):

Reproduction command

# clone and check out commit
git clone https://github.com/allenai/open-instruct.git
git checkout 3f37c29ddc97d2c108a7658692d2d2c3708ef182

# run my exact command for launching exps
for learning_rate in 5e-7; do
for beta in 0.01; do
for nspp in 16; do
for m in half-m ; do
for kl_estimator in kl3; do
local_rollout_batch_size=8
# `half-m` is the same as setting number of mini-batches to be 2.
if [ $m == "half-m" ]; then
    local_mini_batch_size=$(($local_rollout_batch_size * $nspp / 2))
else
    local_mini_batch_size=$(($local_rollout_batch_size * $nspp))
fi
exp_name="0204_lr_scan_grpo_math_lr_${learning_rate}_${kl_estimator}_${beta}_${nspp}_${m}_${RANDOM}"
echo $exp_name:
echo --- local_mini_batch_size=$local_mini_batch_size
echo --- num_gradient_updates=$(($local_rollout_batch_size * $nspp / $local_mini_batch_size))
python open_instruct/grpo_vllm_thread_ray_gtrl.py \
    --exp_name $exp_name \
    --beta $beta \
    --local_mini_batch_size $local_mini_batch_size \
    --number_samples_per_prompt $nspp \
    --output_dir output/$exp_name \
    --local_rollout_batch_size $local_rollout_batch_size \
    --kl_estimator $kl_estimator \
    --learning_rate $learning_rate \
    --dataset_mixer_list allenai/RLVR-GSM-MATH-IF-Mixed-Constraints 1.0 \
    --dataset_mixer_list_splits train \
    --dataset_mixer_eval_list allenai/RLVR-GSM-MATH-IF-Mixed-Constraints 16 \
    --dataset_mixer_eval_list_splits train \
    --max_token_length 2048 \
    --max_prompt_token_length 2048 \
    --response_length 2048 \
    --model_name_or_path allenai/Llama-3.1-Tulu-3-8B-DPO \
    --non_stop_penalty \
    --stop_token eos \
    --temperature 1.0 \
    --ground_truths_key ground_truth \
    --chat_template_name tulu \
    --sft_messages_key messages \
    --total_episodes 10000000 \
    --penalty_reward_value 0.0 \
    --deepspeed_stage 2 \
    --per_device_train_batch_size 2 \
    --local_rollout_forward_batch_size 2 \
    --actor_num_gpus_per_node 6 \
    --num_epochs 1 \
    --vllm_tensor_parallel_size 2 \
    --lr_scheduler_type constant \
    --apply_verifiable_reward true \
    --seed 1 \
    --num_evals 30 \
    --save_freq 40 \
    --reward_model_multiplier 0.0 \
    --gradient_checkpointing \
    --with_tracking
done
done
done
done
done

License and use

All Llama 3.1 Tülu3 models are released under Meta's Llama 3.1 Community License Agreement. Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. Tülu3 is intended for research and educational use. For more information, please see our Responsible Use Guidelines.

The models have been fine-tuned using a dataset mix with outputs generated from third party models and are subject to additional terms: Gemma Terms of Use and Qwen License Agreement (models were improved using Qwen 2.5).

Citation

If Tülu3 or any of the related materials were helpful to your work, please cite:

@article{lambert2024tulu3,
  title = {Tülu 3: Pushing Frontiers in Open Language Model Post-Training},
  author = {
    Nathan Lambert and 
    Jacob Morrison and 
    Valentina Pyatkin and 
    Shengyi Huang and 
    Hamish Ivison and 
    Faeze Brahman and 
    Lester James V. Miranda and 
    Alisa Liu and 
    Nouha Dziri and 
    Shane Lyu and 
    Yuling Gu and 
    Saumya Malik and 
    Victoria Graf and 
    Jena D. Hwang and 
    Jiangjiang Yang and
    Ronan Le Bras and
    Oyvind Tafjord and
    Chris Wilhelm and
    Luca Soldaini and 
    Noah A. Smith and 
    Yizhong Wang and 
    Pradeep Dasigi and 
    Hannaneh Hajishirzi
  },
  year = {2024},
  email = {tulu@allenai.org}
}