shanearora/i-am-a-good-big-instruct-model

OLMo-2-1124-13B-Instruct

OLMo-2 13B Instruct November 2024 is post-trained variant of the OLMo-2 13B November 2024 model, which has undergone supervised finetuning on an OLMo-specific variant of the Tülu 3 dataset and further DPO training on this dataset, and finally RLVR training using this data. 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. Check out the OLMo 2 paper (forthcoming) or Tülu 3 paper for more details!

OLMo is a series of Open Language Models designed to enable the science of language models. These models are trained on the Dolma dataset. We are releasing all code, checkpoints, logs (coming soon), and associated training details. The core models released in this batch include the following:

Stage	OLMo-2 7B	OLMo 2 13B
Base Model	allenai/OLMo2-7B-1124	allenai/OLMo-2-13B-1124
SFT	allenai/OLMo-2-1124-7B-SFT	allenai/OLMo-2-1124-13B-SFT
DPO	allenai/OLMo-2-1124-7B-DPO	allenai/OLMo-2-1124-13B-DPO
Final Models (RLVR)	allenai/OLMo-2-1124-7B-Instruct	allenai/OLMo-2-1124-13B-Instruct
Reward Model (RM)	allenai/OLMo-2-1124-7B-RM	(Same as 8B)

Model description

• Model type: A model trained on a mix of publicly available, synthetic and human-created datasets.
• Language(s) (NLP): Primarily English
• License: Apache 2.0
• Finetuned from model: allenai/OLMo-2-13B-1124-DPO

Model Sources

• Project Page: https://allenai.org/olmo
• Repositories:
  • Core repo (training, inference, fine-tuning etc.): https://github.com/allenai/OLMo
  • Evaluation code: https://github.com/allenai/olmes
  • Further fine-tuning code: https://github.com/allenai/open-instruct
• Paper: Coming soon!
• Demo: https://playground.allenai.org/

Using the model

Loading with HuggingFace

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

from transformers import AutoModelForCausalLM

olmo_model = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-1124-13B-Instruct")

Chat template

The chat template for our models is formatted as:

<|endoftext|><|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:

<|endoftext|><|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 OLMo 2, 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 OLMo 2 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). See the Falcon 180B model card for an example of this.

Performance

Model	Average	AlpacaEval	BBH	DROP	GSM8k	IFEval	MATH	MMLU	Safety	PopQA	TruthQA
Open weights models
Gemma-2-9B-it	51.9	43.7	2.5	58.8	79.7	69.9	29.8	69.1	75.5	28.3	61.4
Ministral-8B-Instruct	52.1	31.4	56.2	56.2	80.0	56.4	40.0	68.5	56.2	20.2	55.5
Mistral-Nemo-Instruct-2407	51.1	45.8	56.0	23.6	81.4	64.5	31.9	70.0	52.7	26.9	57.7
Qwen-2.5-7B-Instruct	57.1	29.7	25.3	54.4	83.8	74.7	69.9	76.6	75.0	18.1	63.1
Llama-3.1-8B-Instruct	58.9	25.8	69.7	61.7	83.4	80.6	42.5	71.3	70.2	28.4	55.1
Tülu 3 8B	60.4	34.0	66.0	62.6	87.6	82.4	43.7	68.2	75.4	29.1	55.0
Qwen-2.5-14B-Instruct	61.0	34.6	35.4	50.5	83.9	82.4	70.6	81.1	79.3	21.1	70.8
Fully open models
OLMo-7B-Instruct	28.2	5.2	35.3	30.7	14.3	32.2	2.1	46.3	54.0	17.1	44.5
OLMo-7B-0424-Instruct	33.2	8.5	35.2	47.9	23.2	39.2	5.2	48.9	49.3	18.9	55.2
OLMoE-1B-7B-0924-Instruct	35.5	8.5	37.2	34.3	47.2	46.2	8.4	51.6	51.6	20.6	49.1
MAP-Neo-7B-Instruct	42.9	17.6	26.4	48.2	69.4	35.9	31.5	56.5	73.7	18.4	51.6
OLMo-2-7B-SFT	50.0	9.3	50.7	58.2	71.2	68.0	25.1	62.0	82.4	25.0	47.8
OLMo-2-7B-DPO	55.0	29.9	47.0	58.8	82.4	74.5	31.2	63.4	81.5	24.5	57.2
OLMo-2-13B-SFT	55.7	12.0	58.8	71.8	75.7	71.5	31.1	67.3	82.8	29.3	56.2
OLMo-2-13B-DPO	61.0	38.3	58.5	71.9	84.2	80.6	35.0	68.5	80.6	28.9	63.9
OLMo-2-7B-1124–Instruct	55.7	31.0	48.9	58.9	85.2	75.6	31.3	63.9	81.2	24.6	56.3
OLMo-2-13B-1124-Instruct	61.4	37.5	58.4	72.1	87.4	80.4	39.7	68.6	77.5	28.8	63.9

Hyperparameters

PPO settings for RLVR:

• Learning Rate: 4 × 10⁻⁷
• Discount Factor (gamma): 1.0
• General Advantage Estimation (lambda): 0.95
• Mini-batches (N_mb): 1
• PPO Update Iterations (K): 4
• PPO's Clipping Coefficient (epsilon): 0.2
• Value Function Coefficient (c1): 0.1
• Gradient Norm Threshold: 1.0
• Learning Rate Schedule: Linear
• Generation Temperature: 1.0
• Batch Size (effective): 512
• Max Token Length: 2,048
• Max Prompt Token Length: 2,048
• Penalty Reward Value for Responses without an EOS Token: -10.0
• Response Length: 2,048
• Total Episodes: 100,000 (this checkpoint is training step 360)
• KL penalty coefficient (beta): 0.03
• Warm up ratio (omega): 0.0

License and use

OLMo 2 is licensed under the Apache 2.0 license. OLMo 2 is intended for research and educational use. For more information, please see our Responsible Use Guidelines. This model has been fine-tuned using a dataset mix with outputs generated from third party models and are subject to additional terms: Gemma Terms of Use.

Citation

A technical manuscript is forthcoming!