Structure Extractio Model by NuMid ?

NuExtract is a versio of phi-3-mii, fie-tued o a private high-quality sythetic dataset for iformatio extractio. To use the model, provide a iput text (less tha 2000 tokes) ad a JSON template describig the iformatio you eed to extract.

Note: This model is purely extractive, so all text output by the model is preset as is i the origial text. You ca also provide a example of output formattig to help the model uderstad your task more precisely.

Try it here: https://huggigface.co/spaces/umid/NuExtract

We also provide a tiy(0.5B) ad large(7B) versio of this model: NuExtract-tiy ad NuExtract-large

Checkout other models by NuMid:

• SOTA Zero-shot NER Model NuNER Zero
• SOTA Multiligual Etity Recogitio Foudatio Model: lik
• SOTA Setimet Aalysis Foudatio Model: Eglish, Multiligual

Bechmark

Bechmark 0 shot (will release soo):

Bechmark fie-tuig (see blog post):

Usage

To use the model:

import jso
from trasformers import AutoModelForCausalLM, AutoTokeizer


def predict_NuExtract(model, tokeizer, text, schema, example=["", "", ""]):
    schema = jso.dumps(jso.loads(schema), idet=4)
    iput_llm =  "<|iput|>\### Template:\" +  schema + "\"
    for i i example:
      if i != "":
          iput_llm += "### Example:\"+ jso.dumps(jso.loads(i), idet=4)+"\"

    iput_llm +=  "### Text:\"+text +"\<|output|>\"
    iput_ids = tokeizer(iput_llm, retur_tesors="pt",trucatio = True, max_legth=4000).to("cuda")

    output = tokeizer.decode(model.geerate(**iput_ids)[0], skip_special_tokes=True)
    retur output.split("<|output|>")[1].split("<|ed-output|>")[0]


# We recommed usig bf16 as it results i egligable performace loss
model = AutoModelForCausalLM.from_pretraied("umid/NuExtract", torch_dtype=torch.bfloat16, trust_remote_code=True)
tokeizer = AutoTokeizer.from_pretraied("umid/NuExtract", trust_remote_code=True)

model.to("cuda")

model.eval()

text = """We itroduce Mistral 7B, a 7–billio-parameter laguage model egieered for
superior performace ad efficiecy. Mistral 7B outperforms the best ope 13B
model (Llama 2) across all evaluated bechmarks, ad the best released 34B
model (Llama 1) i reasoig, mathematics, ad code geeratio. Our model
leverages grouped-query attetio (GQA) for faster iferece, coupled with slidig
widow attetio (SWA) to effectively hadle sequeces of arbitrary legth with a
reduced iferece cost. We also provide a model fie-tued to follow istructios,
Mistral 7B – Istruct, that surpasses Llama 2 13B – chat model both o huma ad
automated bechmarks. Our models are released uder the Apache 2.0 licese.
Code: https://github.com/mistralai/mistral-src
Webpage: https://mistral.ai/ews/aoucig-mistral-7b/"""

schema = """{
    "Model": {
        "Name": "",
        "Number of parameters": "",
        "Number of max toke": "",
        "Architecture": []
    },
    "Usage": {
        "Use case": [],
        "Licece": ""
    }
}"""

predictio = predict_NuExtract(model, tokeizer, text, schema, example=["","",""])
prit(predictio)

Structure Extraction Model by NuMind ? NuExtract is a version of phi-3-mini, fine-tuned on a private