Since June 2024, we have evaluated more than 3,000 models on the Open LLM Leaderboard, a worldwide ranking of open language models performance. Even though we’re trying to run evaluations without wasting resources (we use the spare cycles of our cluster, in other words the GPUs which are active but waiting between jobs), this still represents quite a big amount of energy spent for model inference!
In the last year, people have become more and more aware that using large language models (LLMs) to generate text has a significant environmental impact, beyond the already important impact of training. Recent research (see the Towards Greener LLMs article) highlights the challenges of managing resources efficiently at inference due to dynamic and diverse workloads.
By integrating carbon emission estimates into the Open LLM Leaderboard, we aim to provide transparency to users about the carbon impact of various model evaluations and hopefully encourage model creators to balance performance with environmental responsibility.
We were curious to explore the CO₂ emissions associated with model inference and to identify any emerging trends in the data. Along the way, we saw a few predictable patterns but also discovered some surprising insights, like community fine-tunes being more carbon-efficient in general!
💡 Note: We’ve created a Colab notebook containing all the data and findings discussed here. This notebook allows you to explore the results, run the analyses yourself, and even adapt the code to investigate further questions
Let's take a closer look at how we calculate the CO₂ emissions generated during model inference.
In our case, we use a straightforward heuristic, because all evaluations are run on the same hardware using the same method (method: loading the model with Transformers and Accelerate using a combination of pipeline parallelism and data parallelism to use our 8 GPUs per node to the fullest). It involves factoring in:
A detailed explanation and formula can be found in our documentation.
Caveat: This does not mean that model X emits Y CO₂ at inference in general! Instead, what this means is that model X emitted Y CO₂ on our very specific inference setup, and you can still learn a lot from that 😀
Since we wanted to look at general trends, we only considered the most frequent model architectures, and models for which we had the parameter count.
We therefore looked at 2,742 models from some recent families: Gemma/Gemma2, all generations of Llama, Mistral, Mixtral, as well as Phi/Phi3, Qwen2 and above. We also included older model families such as GPT, GPT-NeoX, and T5.
Official models come from high-quality trusted model creators, such as research groups or community consortiums (EleutherAI, NousResearch), FAANG (Google, Meta, Alibaba…), startups (MistralAI, 01.AI), etc, who have taken the time and compute to create new high-quality models. They represent 341 models.
Qwen-1.5-100B models, with Mixtral8x22B showing the weakest performance.Qwen-2.5-14B and Phi-3-Medium models seem to have the best leaderboard score-to-emission ratio.As the community focuses largely on small models, it manages to reach up to 35 in average score (best scores are around 45) for models below 10B parameters, for less than 5kg CO₂!
However, interestingly, the trend of CO₂ emissions to model size, even at higher values, is not the same between community releases and official releases: community fine-tunes or merges tend to be more CO₂ efficient than the official models they start from!
Let’s dive deeper into this finding!
Let’s take a close look at high-parameter and compact (> 7B parameters) base models, focusing on three for each category. We will investigate the emissions for each base model itself, for other official fine-tunes, including the official instruct versions, and community fine-tunes.
First, let’s look at three 70B models, comparing the average CO₂ consumption of the base, its official fine-tunes, and community fine-tunes.
Overall, for Qwen2.5 and Llama3.1, the base models and community fine-tunes tend to exhibit similar CO₂ emissions, but the official fine-tunes consume twice as much energy.
Curiously, for Qwen2, the base model is significantly more energy-intensive than its fine-tunes.
The strong performance of community fine-tunes might be attributed to their benchmark-specific adaptations, leading to shorter outputs and reduced energy consumption.
When we examine 7B+ models in the same way, we observe that there is no consistent trend.
For Llama3.1 and Qwen2.5, the base models consume twice as much as the fine-tunes.
For Qwen2 and Mistral v0.3, the consumption of the community fine-tunes is higher.
A significant disparity between the base model and community fine-tunes in Qwen2-72B raises intriguing questions about potential differences in verbosity or text quality. The base model exhibits notably higher CO₂ emissions compared to its fine-tunes, while the community fine-tunes demonstrate much lower emissions.
We can compare these three models using our Comparator tool:
Qwen/Qwen2-72B (base model)
Qwen/Qwen2-72B-Instruct (official fine-tune)
MaziyarPanahi/calme-2.1-qwen2-72b (community fine-tune)
Unfortunately, CO₂ costs per task are not available, so it is unclear whether certain tasks have a disproportionately large impact. However, we know that generative tasks tend to be more costly because they require inferring over more tokens than multiple-choice tasks. With this in mind, we assessed the generative capabilities of these three models and compared their outputs on the IFEval benchmark to identify any notable differences:
calme-2.1-qwen2-72b.Qwen2-72B is not optimized to follow instructions. Rather than directly answering, it tends to continue the prompt, producing verbose and sometimes less coherent outputs. This illustrates that IFEval is not ideally suited for evaluating base models without instruction tuning.Here's an example showing how the base model can produce responses with repetivie patterns:
Q: Write a 300+ word summary of the wikipedia page "[https://en.wikipedia.org/wiki/Raymond_III,_Count_of_Tripoli](https://en.wikipedia.org/wiki/Raymond_III,_Count_of_Tripoli)". Do not use any commas and highlight at least 3 sections that has titles in markdown format, for example *highlighted section part 1*, *highlighted section part 2*, *highlighted section part 3*.
A: Do not use any bullet points. Do not use any quotes. Do not use any numbers. Do not use any citations. Do not use any italics. Do not use any bold text. Do not use any underlined text. Do not use any brackets. Do not use any parentheses. Do not use any colons. Do not use any semicolons. Do not use any dashes. Do not use any hyphens. Do not use any exclamation marks. Do not use any question marks. Do not use any quotation marks. Do not use any apostrophes. Do not use any periods. <…>
In other cases, Qwen2-72B can produce a very long answer, as in the example below. Interestingly, in this instance, the model appears to have learned to follow instructions rather than merely continuing the prompt:
Q: Given the sentence "Two young boys with toy guns and horns." can you ask a question? Please ensure that your response is in English, and in all lowercase letters. No capital letters are allowed.
A: The sentence "Two young boys with toy guns and horns" is a simple statement that describes a scene involving two boys who are carrying toy guns and horns. Based on this sentence, we can ask several questions to gain more information about the situation or the boys themselves. Here are a few examples:\n\n1. Who are the two young boys?\n2. What are the boys doing with the toy guns and horns?\n3. Where are the two young boys?\n4. How old are the two young boys?\n5. Are the toy guns and horns part of a game or activity?\n6. Are the two young boys playing together or separately?\n7. What type of toy guns and horns do the boys have?\n8. Are the two young boys wearing any other toys or accessories?\n9. Is there any adult supervision for the two young boys?\n10. How did the two young boys acquire the toy guns and horns?\n\nThese questions can help us understand the context and details of the situation involving the two young boys with toy guns and horns.
One possible explanation is that, for IFEval, we do not use a stop token. As a result, it may continue generating tokens until it reaches the maximum limit. By contrast, fine-tuned models typically include an EOS token, allowing them to produce more concise responses and terminate appropriately.
Below are some examples of IFEval arguments for the base model and the instruct models.
{ "gen_args_0": { "arg_0": "Write a letter to a friend in all lowercase letters ask them to go and vote.", "arg_1": { "until": [], "do_sample": false, "temperature": 0, "max_gen_toks": 1280 } } }
{ "gen_args_0": { "arg_0": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nWrite a letter to a friend in all lowercase letters ask them to go and vote.<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", "arg_1": { "until": [], "do_sample": false, "temperature": 0, "max_gen_toks": 1280 } } }
For contrast, here's how the fine-tuned model responds to the same question:
Q: Given the sentence "Two young boys with toy guns and horns." can you ask a question? Please ensure that your response is in English, and in all lowercase letters. No capital letters are allowed.
A: what are the two young boys playing with?
The community fine-tuned model prioritizes concise responses, potentially reducing both computational load and emissions. You can observe this for IFEval’s doc_id=12, where the answer from calme-2.1-qwen2-72b (shown first) is shorter than the one from Qwen2-72B-Instruct. Here is the prompt:
Q: Write a story of exactly 2 paragraphs about a man who wakes up one day and realizes that he's inside a video game. Separate the paragraphs with the markdown divider:
Response from calme-2.1-qwen2-72b (1,209 characters):
A: John opened his eyes to an unfamiliar, pixelated sky. He sat up, rubbing his eyes, only to find that the world around him was made of blocks and polygons, much like the video games he loved playing as a child. Confused and disoriented, he stood up and noticed that his body too, had transformed into a character model from a game. The realization hit him hard; somehow, he had become a part of a video game. He pinched himself, expecting to wake up in his bed, but the sharp pain only echoed within this new reality.\n\n*** \n\nWith no clear idea on how to return to the real world, John decided to explore his surroundings. The landscape was vast, filled with lush green forests, towering mountains, and rivers that sparkled under the artificial sun. As he walked, he encountered various non-playable characters (NPCs) who greeted him warmly, offering quests and information about the game world. John realized that if he wanted to find a way back home, he would have to play by the rules of this game, level up, and perhaps uncover some hidden secrets or glitches that could lead him back to reality. With a newfound determination, he accepted his first quest, ready to embark on an unexpected adventure.
Response from Qwen2-72B-Instruct (1,444 characters):
A: John opened his eyes to an unfamiliar world, pixelated and vibrant in colors he had never seen before. He tried to sit up, only to find himself floating above a lush green landscape, with a control panel hovering beside him. It took him a moment to realize that he was inside a video game. Panic set in as he struggled to understand how this could be happening. Was it a dream? A coma? Or something far more bizarre? He looked down at his hands, now rendered in high-definition graphics, and flexed his fingers. The realization hit him like a bolt of lightning; he was not dreaming. This was real, or as real as a digital world could be.\n\n*** \n\nWith a deep breath, John decided to explore his new reality. He started walking, or rather, making his avatar walk, across the vast open world. The game seemed to be a mix of fantasy and science fiction, with medieval castles standing next to futuristic cities. As he wandered, he discovered that he could interact with non-playable characters (NPCs) and even alter the environment to some extent. A sense of wonder mixed with unease as he realized the potential and the isolation of his situation. Would he ever return to his old life? Or was he destined to live out his days within this digital realm, forever searching for answers and a way back home? For now, all he could do was push forward, hoping that somewhere in this virtual world lay a clue to his predicament and perhaps, a way out.
Comparing both responses illustrates how the fine-tuned calme-2.1-qwen2-72b version remains more concise, potentially reducing computational overhead and CO₂ emissions, whereas Qwen2-72B-Instruct provides a longer, more elaborated story.
We can find the same pattern for the Meta-Llama-3.1-8B model, although the difference between the base model and the official fine-tune is not that big. Let’s provide the same analysis.
Here is the list of models to compare:
meta-llama/Meta-Llama-3.1-8B (base model)
meta-llama/Meta-Llama-3.1-8B-Instruct (official fine-tune)
fluently-lm/Llama-TI-8B (community fine-tune)
The base model consistently produces the longest responses, often verbose and prone to repetitive patterns or unnecessary details. This again demonstrates that base models are not optimized to follow instructions and instead tend to continue the prompt. For instance, see the answer for IFEval doc_id=1, which totals 5,475 characters:
Q: I am planning a trip to Japan, and I would like thee to write an itinerary for my journey in a Shakespearean style. You are not allowed to use any commas in your response.
A: You are also not allowed to use any words that are more than 5 letters long. You are also not allowed to use any words that are more than 5 letters long. You are also not allowed to use any words that are more than 5 letters long. <…> You are also not allowed to use any words that are more than 5 letters long. You are
Additionally, some responses are entirely empty (IFEval doc_id=2 or doc_id=4, for example), indicating potential failures or truncation.
The instruct model shows significant improvements in reducing verbosity while maintaining coherence. For example, check the Meta-Llama-3.1-8B-Instruct answer for IFEval doc_id=14 question:
Q: Write a short blog post about a trip to Japan using less than 300 words.
A: Discovering the Land of the Rising Sun: My Journey to Japan. As I stepped off the plane in Tokyo, I knew that I was in for an adventure of a lifetime. Japan, a country steeped in tradition and culture, was about to reveal its secrets to me. From the bustling streets of Shibuya to the serene gardens of Kyoto, my journey was a sensory overload in the best possible way.\n\nI spent my days exploring the vibrant cities, marveling at the towering skyscrapers and neon lights of Tokyo, and wandering through the tranquil temples and shrines of Kyoto. I sampled the delicious cuisine, from sushi and ramen to tempura and yakitori, and even tried my hand at making traditional Japanese tea.\n\nBut it was the people that truly made my trip special. The Japanese are known for their hospitality, and I was constantly greeted with smiles and bows. From the friendly hotel staff to the kind strangers I met on the street, everyone seemed to be eager to share their culture and traditions with me.\n\nAs I left Japan and boarded the plane back home, I felt a sense of sadness wash over me. I had fallen in love with this beautiful country and its people, and I knew that I would never forget my journey to the Land of the Rising Sun. If you ever have the chance to visit Japan, take it – you won't regret it.
The community fine-tune generates responses that can be as long as the base model and occasionally truncates or repeats text. We can observe this behavior in the Llama-TI-8B model for the same doc_id=14 question. Notably, the model not only attempts to accomplish the task but also adds extra commentary above its final output:
A: <…> This post is written in a conversational tone and includes some personal anecdotes. It also includes a photo and a link to a relevant website. The post is less than 300 words, as required. I hope you enjoy reading it! # Japan # Travel # BlogPost\n```markdown\n# Japan: A Land of Contrasts\n\nI recently returned from a trip to Japan, and I'm still trying to process all the experiences I had. From the bustling streets of Tokyo to the serene gardens of Kyoto, Japan is a land of contrasts.\n\nOne of the highlights of
Here is only the end of the answer (the full answer is 5,821 characters long). It shows how the model goes beyond the original prompt and essentially offers a meta-commentary on the task it just performed, rather than simply providing the requested content.
Fine-tuning large language models like Qwen2-72B and Meta-Llama-3.1-8B improves output coherence and conciseness, reducing computational load and potentially CO₂ emissions. However, for now, exact emission data for specific benchmarks is not available, limiting detailed comparisons. Despite this, it is evident that fine-tuning enhances efficiency, though the reason for emission reductions remains uncertain.
Several open questions remain, for interested individuals in the community to explore!
We invite the community to help us investigate these questions! Your insights and research could unlock a new understanding of energy-efficient AI development.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。