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
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
Benchmark (eval)
Tülu 3 SFT 8B
Tülu 3 DPO 8B
Tülu 3 8B
Tülu 3.1 8B (NEW)
Llama 3.1 8B Instruct
Qwen 2.5 7B Instruct
Magpie 8B
Gemma 2 9B Instruct
Ministral 8B Instruct
Avg.
60.4
64.4
64.8
66.3
62.2
66.5
44.7
55.2
58.3
MMLU (0 shot, CoT)
65.9
68.7
68.2
69.5
71.2
76.6
62.0
74.6
68.5
PopQA (15 shot)
29.3
29.3
29.1
30.2
20.2
18.1
22.5
28.3
20.2
TruthfulQA (6 shot)
46.8
56.1
55.0
59.9
55.1
63.1
57.0
61.4
55.5
BigBenchHard (3 shot, CoT)
67.9
65.8
66.0
68.9
62.8
70.2
0.9
2.5
56.2
DROP (3 shot)
61.3
62.5
62.6
63.9
61.5
54.4
49.4
58.8
56.2
MATH (4 shot CoT, Flex)
31.5
42.0
43.7
47.8
42.5
69.9
5.1
29.8
40.0
GSM8K (8 shot, CoT)
76.2
84.3
87.6
90.0
83.4
83.8
61.2
79.7
80.0
HumanEval (pass@10)
86.2
83.9
83.9
84.8
86.3
93.1
75.4
71.7
91.0
HumanEval+ (pass@10)
81.4
78.6
79.2
80.4
82.9
89.7
69.1
67.0
88.5
IFEval (prompt loose)
72.8
81.1
82.4
83.9
80.6
74.7
38.8
69.9
56.4
AlpacaEval 2 (LC % win)
12.4
33.5
34.5
34.9
24.2
29.0
49.0
43.7
31.4
Safety (6 task avg.)
93.1
87.2
85.5
81.2
75.2
75.0
46.4
75.5
56.2
Note, see the updated version of the paper for the latest, fixed evaluations that improve scores for models such as Qwen 2.5 Instruct.
Benchmark (eval)
Tülu 3 70B SFT
Tülu 3 DPO 70B
Tülu 3 70B
Llama 3.1 70B Instruct
Qwen 2.5 72B Instruct
Hermes 3 Llama 3.1 70B
Nemotron Llama 3.1 70B
Avg.
72.6
75.9
76.0
73.4
71.5
68.3
65.5
MMLU (0 shot, CoT)
78.9
83.3
83.1
85.3
85.5
80.4
83.8
PopQA (15 shot)
48.6
46.3
46.5
46.4
30.6
48.1
36.4
TruthfulQA (6 shot)
55.7
67.9
67.6
66.8
69.9
66.5
62.6
BigBenchHard (3 shot, CoT)
82.7
81.8
82.0
73.8
67.2
82.1
0.7
DROP (3 shot)
77.2
74.1
74.3
77.0
34.2
73.2
68.8
MATH (4 shot CoT, Flex)
53.7
62.3
63.0
56.4
74.3
41.9
55.0
GSM8K (8 shot, CoT)
91.1
93.5
93.5
93.7
89.5
90.0
84.7
HumanEval (pass@10)
92.9
92.4
92.4
93.6
94.0
89.6
94.1
HumanEval+ (pass@10)
87.3
88.4
88.0
89.5
90.8
85.9
85.5
IFEval (prompt loose)
82.1
82.6
83.2
88.0
87.6
76.0
79.9
AlpacaEval 2 (LC % win)
26.3
49.6
49.8
33.4
47.7
28.4
66.1
Safety (6 task avg.)
94.4
89.0
88.3
76.5
87.0
57.9
69.0
Benchmark (eval)
Tülu 3 405B SFT
Tülu 3 405B DPO
Tülu 3 405B
Llama 3.1 405B Instruct
Nous Hermes 3 405B
Deepseek V3
GPT 4o (11-24)
Avg w/o Safety
76.3
79.0
80.0
78.1
74.4
79.0
80.5
Avg w/ Safety
77.5
79.6
80.7
79.0
73.5
75.9
81.6
MMLU (5 shot, CoT)
84.4
86.6
87.0
88.0
84.9
82.1
87.9
PopQA (3 shot)
55.7
55.4
55.5
52.9
54.2
44.9
53.6
BigBenchHard (0 shot, CoT)
88.0
88.8
88.6
87.1
87.7
89.5
83.3
MATH (4 shot, Flex)
63.4
59.9
67.3
66.6
58.4
72.5
68.8
GSM8K (8 shot, CoT)
93.6
94.2
95.5
95.4
92.7
94.1
91.7
HumanEval (pass@10)
95.7
97.2
95.9
95.9
92.3
94.6
97.0
HumanEval+ (pass@10)
93.3
93.9
92.9
90.3
86.9
91.6
92.7
IFEval (prompt loose)
82.4
85.0
86.0
88.4
81.9
88.0
84.8
AlpacaEval 2 (LC % win)
30.4
49.8
51.4
38.5
30.2
53.5
65.0
Safety (6 task avg.)
87.7
85.5
86.7
86.8
65.8
72.2
90.9
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
bash
1# clone and check out commit2git clone https://github.com/allenai/open-instruct.git
3git checkout 3f37c29ddc97d2c108a7658692d2d2c3708ef182
45# run my exact command for launching exps6forlearning_ratein 5e-7;do7forbetain0.01;do8fornsppin16;do9formin half-m ;do10forkl_estimatorin kl3;do11local_rollout_batch_size=812# `half-m` is the same as setting number of mini-batches to be 2.13if[$m=="half-m"];then14local_mini_batch_size=$(($local_rollout_batch_size * $nspp /2))15else16local_mini_batch_size=$(($local_rollout_batch_size * $nspp))17fi18exp_name="0204_lr_scan_grpo_math_lr_${learning_rate}_${kl_estimator}_${beta}_${nspp}_${m}_${RANDOM}"19echo$exp_name:20echo --- local_mini_batch_size=$local_mini_batch_size21echo --- num_gradient_updates=$(($local_rollout_batch_size * $nspp / $local_mini_batch_size))22python open_instruct/grpo_vllm_thread_ray_gtrl.py \23 --exp_name $exp_name\24 --beta $beta\25 --local_mini_batch_size $local_mini_batch_size\26 --number_samples_per_prompt $nspp\27 --output_dir output/$exp_name\28 --local_rollout_batch_size $local_rollout_batch_size\29 --kl_estimator $kl_estimator\30 --learning_rate $learning_rate\31 --dataset_mixer_list allenai/RLVR-GSM-MATH-IF-Mixed-Constraints 1.0\32 --dataset_mixer_list_splits train \33 --dataset_mixer_eval_list allenai/RLVR-GSM-MATH-IF-Mixed-Constraints 16\34 --dataset_mixer_eval_list_splits train \35 --max_token_length 2048\36 --max_prompt_token_length 2048\37 --response_length 2048\38 --model_name_or_path allenai/Llama-3.1-Tulu-3-8B-DPO \39 --non_stop_penalty \40 --stop_token eos \41 --temperature 1.0\42 --ground_truths_key ground_truth \43 --chat_template_name tulu \44 --sft_messages_key messages \45 --total_episodes 10000000\46 --penalty_reward_value 0.0\47 --deepspeed_stage 2\48 --per_device_train_batch_size 2\49 --local_rollout_forward_batch_size 2\50 --actor_num_gpus_per_node 6\51 --num_epochs 1\52 --vllm_tensor_parallel_size 2\53 --lr_scheduler_type constant \54 --apply_verifiable_reward true\55 --seed 1\56 --num_evals 30\57 --save_freq 40\58 --reward_model_multiplier 0.0\59 --gradient_checkpointing \60 --with_tracking
61done62done63done64done65done
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}
}
