Leaderboard View
Get the current leaderboard rankings for the given leaderboard
A successful response will include a list of JSON objects in the body.
For the LLM leaderboard:
```json
[
{
"id": "str" (submission uuid),
"rank": "int",
"model_name": "str",
"model_size_billions": "optional float",
"source_url": "optional str",
"organization_name": "optional str",
"organization_url": "optional str",
"domain_knowledge_score": "float",
"quantity_extraction_score": "float",
"program_synthesis_score": "float",
"overall_score": "float",
"is_open_source": "bool",
"is_highlighted": "bool",
}
]
```
For the RAG leaderboard:
```json
[
{
"id": "str" (submission uuid),
"rank": "int",
"model_name": "str",
"rag_name: "optional str",
"source_url": "optional str",
"organization_name": "optional str",
"organization_url": "optional str",
"long_document_score": "float",
"overall_score": "float",
"open_source_status": "str",
"is_highlighted": "bool",
}
]
```
GET /api/v1/leaderboard/llm
[ { "id": "5f86b665-e5f4-45ee-8fc4-881c89cb20e1", "model_name": "o1-preview", "source_url": null, "organization_name": "OpenAI", "organization_url": "https://www.openai.com", "is_highlighted": false, "quantity_extraction_score": 0.9237668161434978, "domain_knowledge_score": 0.8931297709923665, "program_synthesis_score": 0.9146341463414634, "overall_score": 0.9105102444924426, "model_size_billions": null, "is_open_source": false, "rank": 1 }, { "id": "0a34287c-6963-4f80-a463-61b92e5f2701", "model_name": "Claude 3.5 Sonnet", "source_url": null, "organization_name": "Anthropic", "organization_url": "https://www.anthropic.com/", "is_highlighted": false, "quantity_extraction_score": 0.9417040358744395, "domain_knowledge_score": 0.8549618320610687, "program_synthesis_score": 0.8455284552845529, "overall_score": 0.8807314410733538, "model_size_billions": null, "is_open_source": false, "rank": 2 }, { "id": "714f578f-027b-417c-a30e-597164356d77", "model_name": "GPT-4o", "source_url": null, "organization_name": "OpenAI", "organization_url": "https://www.openai.com/", "is_highlighted": false, "program_synthesis_score": 0.8617886178861789, "quantity_extraction_score": 0.9372197309417041, "domain_knowledge_score": 0.8396946564885496, "overall_score": 0.8795676684388108, "model_size_billions": null, "is_open_source": false, "rank": 3 }, { "id": "38d0d75d-4222-4bc3-bccc-1c1bc09b26e3", "model_name": "GPT-4 Turbo", "source_url": null, "organization_name": "OpenAI", "organization_url": "https://openai.com/", "is_highlighted": false, "quantity_extraction_score": 0.9551569506726457, "program_synthesis_score": 0.8577235772357723, "domain_knowledge_score": 0.816793893129771, "overall_score": 0.876558140346063, "model_size_billions": null, "is_open_source": false, "rank": 4 }, { "id": "60215c51-46b5-48ad-ae90-cb38229eb807", "model_name": "GPT-4", "source_url": null, "organization_name": "OpenAI", "organization_url": "https://openai.com/", "is_highlighted": false, "program_synthesis_score": 0.7967479674796748, "quantity_extraction_score": 0.9641255605381166, "domain_knowledge_score": 0.8015267175572519, "overall_score": 0.8541334151916811, "model_size_billions": null, "is_open_source": false, "rank": 5 }, { "id": "f9adc042-6aae-4d4c-aee8-7090c4fb0ba0", "model_name": "Mistral Large 2", "source_url": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2407", "organization_name": "Mistral AI", "organization_url": "https://mistral.ai/", "is_highlighted": false, "domain_knowledge_score": 0.7862595419847328, "quantity_extraction_score": 0.9641255605381166, "program_synthesis_score": 0.8089430894308943, "overall_score": 0.8531093973179146, "model_size_billions": 123.0, "is_open_source": true, "rank": 6 }, { "id": "841e9b94-d70b-4dba-9fef-8650c8017a2d", "model_name": "Claude 3 Opus", "source_url": null, "organization_name": "Anthropic", "organization_url": "https://www.anthropic.com/", "is_highlighted": false, "domain_knowledge_score": 0.7404580152671756, "program_synthesis_score": 0.8252032520325203, "quantity_extraction_score": 0.9282511210762332, "overall_score": 0.8313041294586431, "model_size_billions": null, "is_open_source": false, "rank": 7 }, { "id": "3e67cd57-222b-487e-8ccd-92cd685c6312", "model_name": "Gemini 1.5 Pro", "source_url": null, "organization_name": "Google", "organization_url": "https://www.google.com", "is_highlighted": false, "program_synthesis_score": 0.7357723577235772, "quantity_extraction_score": 0.9551569506726457, "domain_knowledge_score": 0.7709923664122137, "overall_score": 0.8206405582694788, "model_size_billions": null, "is_open_source": false, "rank": 8 }, { "id": "280ab364-27a1-4def-99e9-0679d2f1d6f3", "model_name": "Claude 3 Sonnet", "source_url": null, "organization_name": "Anthropic", "organization_url": "https://www.anthropic.com/", "is_highlighted": false, "program_synthesis_score": 0.7113821138211383, "quantity_extraction_score": 0.9551569506726457, "domain_knowledge_score": 0.7175572519083969, "overall_score": 0.7946987721340603, "model_size_billions": null, "is_open_source": false, "rank": 9 }, { "id": "a9534f00-9871-41fb-af99-b1702c41187d", "model_name": "Llama 3 70B", "source_url": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct", "organization_name": "Meta", "organization_url": "https://facebook.com", "is_highlighted": false, "domain_knowledge_score": 0.7709923664122137, "quantity_extraction_score": 0.9327354260089686, "program_synthesis_score": 0.6788617886178862, "overall_score": 0.7941965270130228, "model_size_billions": 70.0, "is_open_source": true, "rank": 10 }, { "id": "b4078ddc-1c2d-4e6f-80bc-cd7901ca0e60", "model_name": "Mistral Large", "source_url": null, "organization_name": "Mistral AI", "organization_url": "https://mistral.ai/", "is_highlighted": false, "program_synthesis_score": 0.6991869918699187, "domain_knowledge_score": 0.6183206106870229, "quantity_extraction_score": 0.9282511210762332, "overall_score": 0.7485862412110583, "model_size_billions": null, "is_open_source": false, "rank": 11 }, { "id": "07380dd1-3dec-4df8-a560-6db166a9a518", "model_name": "Claude 2", "source_url": null, "organization_name": "Anthropic", "organization_url": "https://www.anthropic.com/", "is_highlighted": false, "quantity_extraction_score": 0.8699551569506726, "program_synthesis_score": 0.6666666666666666, "domain_knowledge_score": 0.6870229007633588, "overall_score": 0.7412149081268993, "model_size_billions": null, "is_open_source": false, "rank": 12 }, { "id": "44603848-2675-47dc-9203-0fa0f36c4d77", "model_name": "Mistral Medium", "source_url": null, "organization_name": "Mistral AI", "organization_url": "https://mistral.ai/", "is_highlighted": false, "program_synthesis_score": 0.6666666666666666, "quantity_extraction_score": 0.8968609865470852, "domain_knowledge_score": 0.6412213740458015, "overall_score": 0.7349163424198512, "model_size_billions": null, "is_open_source": false, "rank": 13 }, { "id": "32aa9bfb-76b3-460f-9732-b7da902c98d9", "model_name": "GPT-3.5 Turbo", "source_url": null, "organization_name": "OpenAI", "organization_url": "https://openai.com/", "is_highlighted": false, "quantity_extraction_score": 0.9237668161434978, "domain_knowledge_score": 0.6030534351145038, "program_synthesis_score": 0.6747967479674797, "overall_score": 0.7338723330751605, "model_size_billions": null, "is_open_source": false, "rank": 14 }, { "id": "f002d245-f7e4-42b1-b72e-9f4609420801", "model_name": "Gemini Pro", "source_url": null, "organization_name": "Google", "organization_url": "https://google.com", "is_highlighted": false, "domain_knowledge_score": 0.5648854961832062, "program_synthesis_score": 0.6341463414634146, "quantity_extraction_score": 0.9417040358744395, "overall_score": 0.71357862450702, "model_size_billions": null, "is_open_source": false, "rank": 15 }, { "id": "8bdb8786-01e6-4663-8114-4fb168f21370", "model_name": "Mistral Small", "source_url": null, "organization_name": "Mistral AI", "organization_url": "https://mistral.ai/", "is_highlighted": false, "quantity_extraction_score": 0.9103139013452914, "domain_knowledge_score": 0.5801526717557252, "program_synthesis_score": 0.6260162601626016, "overall_score": 0.7054942777545393, "model_size_billions": null, "is_open_source": false, "rank": 16 }, { "id": "ad21089d-6cc8-4874-92b6-80b72c547832", "model_name": "Claude 3 Haiku", "source_url": null, "organization_name": "Anthropic", "organization_url": "https://www.anthropic.com/", "is_highlighted": false, "domain_knowledge_score": 0.5725190839694656, "program_synthesis_score": 0.6504065040650406, "quantity_extraction_score": 0.8430493273542601, "overall_score": 0.6886583051295888, "model_size_billions": null, "is_open_source": false, "rank": 17 }, { "id": "4a6a8051-707a-4620-9871-f806586b1b99", "model_name": "Llama 3 8B", "source_url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct", "organization_name": "Meta", "organization_url": "https://facebook.com", "is_highlighted": false, "quantity_extraction_score": 0.8565022421524664, "domain_knowledge_score": 0.6030534351145038, "program_synthesis_score": 0.45934959349593496, "overall_score": 0.6396350902543017, "model_size_billions": 8.0, "is_open_source": true, "rank": 18 }, { "id": "dd8152cc-44d1-454c-b4f0-19372627c201", "model_name": "Llama 2 70B", "source_url": "https://huggingface.co/meta-llama/Llama-2-70b-chat-hf", "organization_name": "Meta", "organization_url": "https://facebook.com", "is_highlighted": false, "quantity_extraction_score": 0.8923766816143498, "program_synthesis_score": 0.4268292682926829, "domain_knowledge_score": 0.48091603053435117, "overall_score": 0.600040660147128, "model_size_billions": 70.0, "is_open_source": true, "rank": 19 }, { "id": "99333ec8-afa1-477b-bf5a-17e01299451e", "model_name": "Zephyr 7B Beta", "source_url": "https://huggingface.co/HuggingFaceH4/zephyr-7b-beta", "organization_name": "Hugging Face", "organization_url": "https://huggingface.co", "is_highlighted": false, "domain_knowledge_score": 0.48854961832061067, "program_synthesis_score": 0.43089430894308944, "quantity_extraction_score": 0.8071748878923767, "overall_score": 0.5755396050520255, "model_size_billions": 7.0, "is_open_source": true, "rank": 20 }, { "id": "17d9609d-dd35-447b-b85a-bedec6af6f54", "model_name": "Mistral Tiny", "source_url": null, "organization_name": "Mistral AI", "organization_url": "https://mistral.ai/", "is_highlighted": false, "program_synthesis_score": 0.34959349593495936, "quantity_extraction_score": 0.8251121076233184, "domain_knowledge_score": 0.5419847328244275, "overall_score": 0.5722301121275685, "model_size_billions": null, "is_open_source": false, "rank": 21 }, { "id": "ead5d804-2ac0-4192-a541-5d93bfc49ee1", "model_name": "Phi 3 Mini 4k", "source_url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", "organization_name": "Microsoft", "organization_url": "https://microsoft.com", "is_highlighted": false, "quantity_extraction_score": 0.6591928251121076, "domain_knowledge_score": 0.5725190839694656, "program_synthesis_score": 0.3821138211382114, "overall_score": 0.5379419100732615, "model_size_billions": 3.8, "is_open_source": true, "rank": 22 }, { "id": "eda94370-f2d3-4474-ab53-ccf4d9b2f169", "model_name": "Starling-7B", "source_url": "https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha", "organization_name": "UC Berkeley", "organization_url": "https://starling.cs.berkeley.edu/", "is_highlighted": false, "domain_knowledge_score": 0.46564885496183206, "quantity_extraction_score": 0.757847533632287, "program_synthesis_score": 0.34552845528455284, "overall_score": 0.5230082812928906, "model_size_billions": 7.0, "is_open_source": true, "rank": 23 }, { "id": "95a0a1a5-3010-475b-a971-75702541c228", "model_name": "Gemma 7B", "source_url": "https://huggingface.co/google/gemma-7b-it", "organization_name": "Google", "organization_url": "https://google.com", "is_highlighted": false, "program_synthesis_score": 0.37398373983739835, "quantity_extraction_score": 0.8026905829596412, "domain_knowledge_score": 0.37404580152671757, "overall_score": 0.516906708107919, "model_size_billions": 8.5, "is_open_source": true, "rank": 24 } ]