SynthGen

Generate compact-but-sufficient prompt→response pairs by prompting your teacher model (via Ollama) with style instructions, collect, tokenize & filter, save as JSONL (or Arrow), and provide a CLI parameter to target dataset size for 3B / 8B / 14B student models to train a LoRA.

Setup

1. Install the dependencies:

   pip install -r requirements.txt

   (Ensure you have requests, transformers, tokenizers, datasets, fuzzywuzzy, python-Levenshtein, tqdm installed).

2. Make sure you have Ollama running. If using a remote host or different port, set the OLLAMA_URL environment variable or pass --teacher-url.

Usage

Generate Dataset

Generate a dataset targeted to 3B models (~2M tokens):

python generate_dataset.py generate \
  --target 3b \
  --style "Always answer as a high medieval English aristocrat, formal and archaic." \
  --teacher-url "http://localhost:11434" \
  --out-file data/style_3b.jsonl

Generate by explicit token budget:

python generate_dataset.py generate \
  --target-tokens 2000000 \
  --batch-size 8 \
  --out-file data/style_bytokens.jsonl

Show Statistics

Show dataset stats:

python generate_dataset.py stats --in-file data/style_3b.jsonl

Convert to Arrow

Convert the JSONL dataset to Apache Arrow format:

python generate_dataset.py convert \
  --in-file data/style_3b.jsonl \
  --out-file data/style_3b.arrow

Sample a Validation Set

Create a small holdout evaluation set:

python generate_dataset.py sample \
  --in-file data/style_3b.jsonl \
  --n 200