### Install Dependencies

Source: https://github.com/beeevita/evoprompt/blob/main/README.md

Install project dependencies using pip.

```bash
pip install -r requirements.txt
```

--------------------------------

### Get Dataset Verbalizers

Source: https://context7.com/beeevita/evoprompt/llms.txt

This snippet demonstrates how to retrieve verbalizers (label words) for a given dataset using `get_dataset_verbalizers`. It shows examples for 'sst2' and 'sst-5' datasets.

```python
from utils import get_dataset_verbalizers

# Get verbalizers (label words) for a dataset
verbalizers = get_dataset_verbalizers("sst2")
print(verbalizers)  # Output: [' negative', ' positive']

verbalizers = get_dataset_verbalizers("sst-5")
print(verbalizers)  # Output: [' terrible', ' bad', ' okay', ' good', ' great']
```

--------------------------------

### Initialize LLM Client and Query

Source: https://context7.com/beeevita/evoprompt/llms.txt

Initialize the LLM client using configuration from 'auth.yaml' and query for task completion or prompt evolution. Supports batch queries.

```python
from llm_client import llm_init, llm_query, paraphrase

# Initialize LLM configuration from auth.yaml
config = llm_init(
    auth_file="./auth.yaml",
    llm_type="turbo",  # Options: 'davinci', 'turbo', 'gpt4'
    setting="default"
)

# Query LLM for task completion
response = llm_query(
    data="Classify the following text: 'This movie was amazing!'",
    client=None,
    type="turbo",
    task=True,  # True for task execution, False for evolution
    temperature=0,
    **config
)
print(response)  # Output: "positive"
```

```python
# Batch query for multiple inputs
batch_data = [
    "Classify: 'Great product!'",
    "Classify: 'Terrible experience.'"
]
responses = llm_query(
    data=batch_data,
    client=None,
    type="davinci",
    task=True,
    temperature=0,
    **config
)
print(responses)  # Output: ["positive", "negative"]
```

```python
# Generate prompt variations using paraphrase
original_prompt = "Classify the sentiment of the text."
variations = paraphrase(
    sentence=original_prompt,
    client=None,
    type="turbo",
    temperature=0.5,
    **config
)
print(variations)  # Output: "Determine the emotional tone of the given text."
```

```python
# Batch paraphrase for multiple prompts
prompts = [
    "Simplify this sentence.",
    "Make this text easier to read."
]
paraphrased = paraphrase(
    sentence=prompts,
    client=None,
    type="davinci",
    temperature=0.5,
    **config
)
# Output: ["Reword this sentence simply.", "Transform this text for easier comprehension."]
```

--------------------------------

### Initialize and Run Genetic Algorithm Evolution

Source: https://context7.com/beeevita/evoprompt/llms.txt

This snippet demonstrates initializing a Genetic Algorithm (GA) evoluter, initializing its population from prompts, running the evolution process, and accessing the best evolved prompt and its score. It also shows how to switch to Differential Evolution (DE) by changing the `evo_mode` and re-initializing the evoluter.

```python
args = parse_args()
args.dataset = "sst2"
args.task = "cls"
args.evo_mode = "ga"
args.popsize = 10
args.budget = 10
args.sel_mode = "wheel"
args.ga_mode = "topk"

# Initialize evaluator
evaluator = CLSEvaluator(args)

# Initialize GA evoluter
ga_evoluter = GAEvoluter(args, evaluator)

# Initialize population from prompts
evaluated_prompts, cur_budget = ga_evoluter.init_pop()
print(f"Initial population size: {len(ga_evoluter.population)}")
print(f"Best initial prompt: {ga_evoluter.population[0]}")

# Run evolution
ga_evoluter.evolute()

# Access results
print(f"Best evolved prompt: {ga_evoluter.population[0]}")
print(f"Best score: {ga_evoluter.scores[0]}")

# For Differential Evolution
args.evo_mode = "de"
args.template = "v1"
de_evoluter = DEEvoluter(args, evaluator)
de_evoluter.init_pop()
de_evoluter.evolute()

# Results are written to output directory:
# - step{N}_pop.txt: Population state at step N
# - dev_result.txt: Final sorted results
```

--------------------------------

### Select initial population subset

Source: https://context7.com/beeevita/evoprompt/llms.txt

Initializes a population of prompts using either top-k selection or random sampling.

```python
all_prompts = ["prompt1", "prompt2", "prompt3", "prompt4", "prompt5"]
top_prompts = k_init_pop("topk", all_prompts, k=3)
print(top_prompts)  # Output: ["prompt1", "prompt2", "prompt3"]

random_prompts = k_init_pop("randomk", all_prompts, k=3)
print(random_prompts)  # Output: 3 randomly selected prompts
```

--------------------------------

### Initialize Summarization Evaluator

Source: https://context7.com/beeevita/evoprompt/llms.txt

Set up a summarization evaluator to test prompt effectiveness for generating summaries. It handles input data, reference texts, and calculates relevant metrics.

```python
# Initialize summarization evaluator
args.task = "sum"
args.dataset = "sam"
sum_evaluator = SumEvaluator(args)

# Evaluate summarization prompt
result = sum_evaluator.forward(
    prompt_pre="Write a one-sentence summary.",
    eval_src=sum_evaluator.dev_src[:10],
    ref_texts=sum_evaluator.dev_tgt[:10]
)
print(result)
# Output: {"hypos": [...], "scores": [0.35, 0.12, 0.28, 0.25]}  # ROUGE-1, ROUGE-2, ROUGE-L, mean
```

--------------------------------

### Run Genetic Algorithm for Simplification Task (Alpaca)

Source: https://github.com/beeevita/evoprompt/blob/main/README.md

Execute the genetic algorithm for prompt evolution on the Alpaca dataset for simplification tasks. Customize --llm_type for different LLMs.

```bash
bash scripts/sim/run_ga_alpaca.sh
```