gistfile1.txt

## Step 1: Create a Sample from GSM8K

Create a Python script named `prepare_gsm8k.py`:

```python
import pandas as pd
import os
from datasets import load_dataset

# Create directories if they don't exist
os.makedirs('data/output', exist_ok=True)

# Load GSM8K dataset
gsm8k = load_dataset('gsm8k', 'main')

# Take 50 samples from the training set
samples = gsm8k['train'].select(range(50))

# Create a text file with the questions
with open('data/output/gsm8k_sample.txt', 'w') as f:
    for i, item in enumerate(samples):
        f.write(f"Problem {i+1}:\n{item['question']}\n\n")

print(f"Created sample with {len(samples)} problems")
print(f"Sample saved to data/output/gsm8k_sample.txt")
```

Run the script:

```bash
python prepare_gsm8k.py
```

## Step 2: Generate Chain-of-Thought Examples

Export your  API key:

```bash
export LLAMA_API_KEY="your-together-api-key"
```

Run the create command to generate CoT examples:

```bash
synthetic-data-kit create data/output/gsm8k_sample.txt --type cot
```

This will:
- Process the sample GSM8K problems
- Generate detailed chain-of-thought reasoning for each problem
- Save the output to `data/generated/gsm8k_sample_cot_examples.json`

## Step 3: Curate Examples by Quality

Next, curate the generated examples to keep only high-quality reasoning:

```bash
synthetic-data-kit curate data/generated/gsm8k_sample_cot_examples.json --threshold 8.0
```

This will:
- Rate each example for quality on a scale of 1-10
- Filter out examples below the threshold
- Save the curated examples to `data/cleaned/gsm8k_sample_cot_examples_cleaned.json`

## Step 4: Save in Fine-Tuning Format

Finally, save the curated examples in a format suitable for fine-tuning:

```bash
# Save as ChatML format (for Llama models)
synthetic-data-kit save-as data/cleaned/gsm8k_sample_cot_examples_cleaned.json -f chatml

# Alternatively, save as OpenAI fine-tuning format
synthetic-data-kit save-as data/cleaned/gsm8k_sample_cot_examples_cleaned.json -f ft
```

The output will be saved to `data/final/`.

## Step 5: Automate the Entire Pipeline

Create a bash script `run_gsm8k_pipeline.sh`:

```bash
#!/bin/bash

# Set your API key
export LLAMA_API_KEY="your-together-api-key"

# Step 1: Prepare the GSM8K sample
python prepare_gsm8k.py

# Step 2: Generate CoT examples
synthetic-data-kit create data/output/gsm8k_sample.txt --type cot

# Step 3: Curate examples
synthetic-data-kit curate data/generated/gsm8k_sample_cot_examples.json --threshold 8.0

# Step 4: Save in fine-tuning format
synthetic-data-kit save-as data/cleaned/gsm8k_sample_cot_examples_cleaned.json -f chatml

echo "Pipeline complete! Fine-tuning data ready in data/final/"
```

Make it executable and run:

```bash
chmod +x run_gsm8k_pipeline.sh
./run_gsm8k_pipeline.sh
```

## Examining the Generated Examples

The generated CoT examples follow this structure:

```json
{
  "question": "How many clips did Natalia sell in total during April and May?",
  "reasoning": "Step 1: First, I need to determine the number of clips Natalia sold in April, which is given as 48.\nStep 2: Then, I calculate the number of clips she sold in May, which is half of the number sold in April.\nStep 3: To find the total number of clips sold, I add the number of clips sold in April and May.",
  "answer": "72"
}
```

These examples can be used to fine-tune models to develop better chain-of-thought reasoning capabilities for solving mathematical problems.