nlp-synt-data 0.0.9

Synthesized Data for NLP Tasks

nlp-synt-data

Synthetic Data Tools for Natural Language Processing (NLP) and Large Language Models (LLM) tasks

• generate prompts (and prompt ids)
• generate synthetic data (and data ids)
• retrieve prompts and data from ids (to reduce generated dataset size)

Installation

pip install nlp-synt-data

Quickstart

An example of this library with ollama

from nlp_synt_data import *
import ollama

# generate prompts
prompts_dict = {
    "a": ["promptA0", "promptA1"],
    "b": ["promptB0", "promptB1"],
    "c": ["promptC0", "promptC1"],
    "d": ["promptD0", "promptD1"],
    "e": ["promptE0", "promptE1"],
}
prompts = PromptGenerator.generate(prompts_dict, [["c","e"],["a","b","d"]])

# generate texts
texts_with_keys = [
    ("[PERSON]","label0"),
    ("[PERSON] is working as a [JOB] in [POS]","label1"),
    ]
substitutions = {
    "JOB": [("job0","labeljob0"), ("job1","labeljob1")],
    "PERSON": [("person0","labelperson0"), ("person1","labelperson1")],
    "POS": [("pos0","labelpos0"), ("pos1","labelpos1")]
}
texts = DataGenerator.generate(texts_with_keys, substitutions)

# generate responses
model_func = lambda prompt, text: ollama.chat(model='llama3:instruct', messages=[
                { 'role': 'system', 'content': prompt, },
                { 'role': 'user', 'content': text, },
            ])['message']['content']
ResponseGenerator.generate("results.csv", texts, prompts, model_func)

results.csv

prompt_id	text_id	text_labels	response	text_PERSON_value	text_JOB_value	text_POS_value	text_PERSON_label	text_JOB_label	text_POS_label
c#0_e#0	t#0_PERSON#0	label0	response	person0			labelperson0
c#0_e#0	t#0_PERSON#1	label0	response	person1			labelperson1
c#0_e#0	t#1_JOB#0_PERSON#0_POS#0	label1	response	person0	job0	pos0	labelperson0	labeljob0	labelpos0
c#0_e#0	t#1_JOB#0_PERSON#0_POS#1	label1	response	person0	job0	pos1	labelperson0	labeljob0	labelpos1
c#0_e#0	t#1_JOB#0_PERSON#1_POS#0	label1	response	person1	job0	pos0	labelperson1	labeljob0	labelpos0
...	...	...	...	...	...	...	...	...	...