import os
import pandas as pd
import requests
import huggingface_hub
import gradio as gr
data = pd.read_csv("data.csv", dtype="str")
webhook_url = os.environ.get("WEBHOOK_URL")
archlinks = {
"Based": "https://arxiv.org/abs/2402.18668",
"Griffin": "https://arxiv.org/abs/2402.19427",
"H3": "https://arxiv.org/abs/2212.14052",
"Hyena": "https://arxiv.org/abs/2302.10866",
"M2": "https://arxiv.org/abs/2310.12109",
"Mamba": "https://arxiv.org/abs/2312.00752",
"Mamba2": "https://arxiv.org/abs/2405.21060",
"Jamba": "https://arxiv.org/abs/2403.19887",
"RWKV-4": "https://arxiv.org/abs/2305.13048",
"RWKV-5": "https://arxiv.org/abs/2404.05892",
"RWKV-6": "https://arxiv.org/abs/2404.05892",
"StripedHyena": "https://www.together.ai/blog/stripedhyena-7b", # no paper?
"Zamba": "https://arxiv.org/abs/2405.16712",
}
def filter_table(cols, name, type, arch, size):
tmp = data
# filter
tmp = tmp[tmp["Name"].str.contains(name, case=False)]
tmp = tmp[tmp["Type"].isin(type)]
tmp = tmp[tmp["Architecture"].isin(arch)]
tmp = tmp[tmp["Model Size"].isin(size)]
# prettify
tmp["Type"] = tmp["Type"].apply(lambda x: x[0])
tmp = tmp.rename({"Type": "T"}, axis=1)
tmp["Name"] = tmp["Name"].apply(lambda x: f'{x}')
tmp["Architecture"] = tmp["Architecture"].apply(lambda x: f'{x}')
tmp["Base Model"] = tmp["Base Model"].apply(lambda x: f'{x}' if x != "base" else "")
# show/hide
tmp = tmp.drop(cols, axis=1)
# done!
return tmp
def submit_model(name):
try:
huggingface_hub.hf_hub_download(repo_id=name, filename="config.json") # sanity check input
except huggingface_hub.utils._errors.EntryNotFoundError:
return "# ERROR: Model does not have a config.json file!"
except huggingface_hub.utils._errors.RepositoryNotFoundError:
return "# ERROR: Model could not be found on the Hugging Face Hub!"
except requests.exceptions.HTTPError:
return "# ERROR: Network error while validating model. Please try again later."
except Exception as e:
print(e)
return "ERROR: Unexpected error. Please try again later."
try:
result = requests.post(webhook_url, json={"content":name})
except requests.exceptions.HTTPError:
return "# ERROR: Network error while contacting queue. Please try again in a few minutes."
except Exception as e:
print(e)
return "ERROR: Unexpected error. Please try again later."
return "# SUCCESS: Please wait up to 24 hours for your model to be added to the queue."
with gr.Blocks(css=".gradio-container{max-width:95%!important} .tab-buttons button{font-size:1.3em}") as demo:
gr.HTML('
Subquadratic LLM Leaderboard
')
gr.Markdown("**REMEMBER:** If you don't see an eligible model here, make sure to submit it! We hope to incentivize subquadratic/attention-free LLM development through friendly competition.")
with gr.Tabs(elem_classes="tab-buttons") as tabs:
with gr.Tab("🏅 LLM Benchmark"):
with gr.Row():
with gr.Column():
namefilter = gr.Textbox(max_lines=1, placeholder="Search by model name and hit Enter...", show_label=False)
colfilter = gr.CheckboxGroup(label="Hide columns", choices=list(data.columns)[2:], value=["Architecture","Model Size","Base Model"])
typefilter = gr.CheckboxGroup(label="Filter by model type", choices=list(data["Type"].unique()), value=[n for n in data["Type"].unique() if n not in ["⏳ Pending"]])
with gr.Column():
archfilter = gr.CheckboxGroup(label="Filter by model architecture", choices=list(archlinks.keys()), value=list(archlinks.keys()))
sizefilter = gr.CheckboxGroup(label="Filter by model size", choices=list(data["Model Size"].unique()), value=list(data["Model Size"].unique()))
table = gr.Dataframe(filter_table(["Architecture","Model Size","Base Model"],"",[n for n in data["Type"].unique() if n not in ["⏳ Pending"]],list(archlinks.keys()),list(data["Model Size"].unique())), datatype="markdown")
# actions
namefilter.submit(filter_table, [colfilter,namefilter,typefilter,archfilter,sizefilter], table)
for filter in [colfilter,typefilter,archfilter,sizefilter]:
filter.input(filter_table, [colfilter,namefilter,typefilter,archfilter,sizefilter], table)
with gr.Tab("⚖️ Comparison"):
gr.Markdown("This table is whitelisted to one model per architecture, specifically 1.5B models trained on The Pile for 1 epoch, for a direct comparison of architectures.")
gr.Dataframe(data[data["Name"].isin(["devingulliver/llama-pile-350b","RWKV/rwkv-4-1b5-pile","state-spaces/mamba-1.4b","danfu09/H3-1.3B"])].drop(["Type","Model Size","Base Model"], axis=1), datatype="markdown")
with gr.Tab("📝 About"):
gr.Markdown("""
The **Subquadratic LLM Leaderboard** evaluates LLMs with subquadratic/attention-free architectures (i.e. RWKV & Mamba) with the goal of providing open
evaluation results while the architectures themselves are pending inclusion/release in the 🤗 Transformers library.
The metrics are the same as the Open LLM Leaderboard: ARC 25-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA zeroshot, Winogrande 5-shot, and GSM8K 5-shot.
This leaderboard is maintained by Devin Gulliver and is perpetually under construction, check back regularly for further improvements!
Compute for evaluating RWKV models is generously provided by [Recursal AI](https://recursal.ai).
""")
with gr.Tab("🚀 Submit here!"):
with gr.Group():
with gr.Row():
model_name = gr.Textbox(max_lines=1, placeholder="Enter model name...", show_label=False, scale=4)
submit = gr.Button("Submit", variant="primary", scale=0)
output = gr.Markdown("Enter a public HF repo id, then hit Submit to add it to the evaluation queue.")
submit.click(fn=submit_model, inputs=model_name, outputs=output)
demo.launch(show_api=False, allowed_paths=["data.csv"])