phi2-ner-sft-v1

Phi-2 model fine-tuned for named entity recognition task

The model was fine-tuned using one quarter of the ConLL 2012 OntoNotes v5 dataset.

• Dataset Source: conll2012_ontonotesv5
• Subset Used: English_v12
• Number of Examples: 21,817

The prompts and expected outputs were constructed as described in [1].

Example input:

I am an excelent linquist. The task is to label location entities in the given sentence. Below are some examples

Input: Only France and Britain backed Fischler's proposal.
Output: Only @@France## and @@Britain## backed Fischler's proposal.

Input: Germany imported 47,000 sheeps from Britain last year, nearly half of total imports.
Output: @@Germany## imported 47,000 sheeps from @@Britain## last year, nearly half of total imports.

Input: It brought in 4275 tonnes of British mutton, some 10% of overall imports.
Output: It brought in 4275 tonnes of British mutton, some 10% of overall imports.

Input: China says Taiwan spoils atmosphere for talks.
Output: 

Expected output:

@@China## says @@Taiwan## spoils atmosphere for talks.

Model Trained Using AutoTrain

This model was trained using SFT AutoTrain trainer. For more information, please visit AutoTrain.

Hyperparameters:

{
    "model": "pahautelman/phi2-ner",
    "train_split": "train",
    "valid_split": null,
    "add_eos_token": false,
    "block_size": 1024,
    "model_max_length": 1024,
    "padding": null,
    "trainer": "sft",
    "use_flash_attention_2": false,
    "log": "none",
    "disable_gradient_checkpointing": false,
    "logging_steps": -1,
    "evaluation_strategy": "epoch",
    "save_total_limit": 1,
    "save_strategy": "epoch",
    "auto_find_batch_size": false,
    "mixed_precision": "fp16",
    "lr": 0.0002,
    "epochs": 1,
    "batch_size": 1,
    "warmup_ratio": 0.1,
    "gradient_accumulation": 4,
    "optimizer": "adamw_torch",
    "scheduler": "linear",
    "weight_decay": 0.01,
    "max_grad_norm": 1.0,
    "seed": 42,
    "apply_chat_template": false,
    "quantization": "int4",
    "target_modules": null,
    "merge_adapter": false,
    "peft": true,
    "lora_r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "model_ref": null,
    "dpo_beta": 0.1,
}

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "pahautelman/phi2-ner-sft-v1"

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path
).eval()

prompt = 'Label the person entities in the given sentence: Russian President Vladimir Putin is due to arrive in Havana a few hours from now to become the first post-Soviet leader to visit Cuba.'

inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors='pt')
outputs = model.generate(
    inputs.to(model.device),
    max_new_tokens=9,
    do_sample=False,
)
output = tokenizer.batch_decode(outputs)[0]

# Model response: "Output: Russian President, Vladimir Putin"
print(output)

References:

[1] Wang et al., GPT-NER: Named entity recognition via large language models 2023