Skip to content

guess

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

guess

README.md

Guess What - 20 Questions for Z80

A 20 Questions game where a tiny neural network learns to answer YES/NO/MAYBE about a secret topic.

Quick Start

# Generate training data (requires ANTHROPIC_API_KEY for --claude)
./gendata.py -t 'elephant' -n 500 --claude > elephant.txt

# Add distractors and nonsense responses
./gendata.py -t 'elephant' -d 30 --nonsense --claude >> elephant.txt

# Generate YES-only questions (helps balance)
./gendata.py -t 'elephant' --yes-only --claude >> elephant.txt

# Generate WIN guesses (ways to guess the answer)
./gendata.py -t 'elephant' --win-only --claude >> elephant.txt

# Balance, train, and build
cat *.txt | ./balance.py -t 5000 -o --stats | ../../feedme.py
../../buildz80com.py -o GUESS.COM

# Play!
../../cpm GUESS.COM

Training Data Format

is it big|YES
can it fly|NO
does it have a trunk|YES
is it grey|MAYBE
elephant|WIN
hello|IDK

Tools

gendata.py

Generates Q&A pairs using an LLM.

# Local Ollama
./gendata.py --topic elephant -n 100

# Claude API (faster, better quality)
./gendata.py --topic elephant --claude -n 500

# With paraphrases (5 variations per question)
./gendata.py --topic elephant --claude -p 5 -n 500

Options:

  • -t, --topic - The secret topic
  • -n, --num - Number of questions to generate
  • -d, --distractors - Add wrong-topic questions (answered NO)
  • --nonsense - Add off-topic phrases (answered IDK)
  • --claude - Use Claude API instead of Ollama
  • --yes-only - Only generate YES-answer questions
  • --win-only - Generate winning guesses
  • -p, --paraphrase - Generate N paraphrases per question

balance.py

Balances training data by over/undersampling classes.

# Balance to 5000 per class, oversample minorities
cat *.txt | ./balance.py -t 5000 -o --stats > balanced.txt

# Check distribution without balancing
cat *.txt | ./balance.py --stats > /dev/null

analyze.py

Analyzes answer distribution in training data.

./analyze.py training.txt

Limitations

The Z80 model uses trigram hashing with 128 buckets:

  • Short distinctive phrases work best
  • Long similar phrases may collide ("bigger than X" vs "smaller than X")
  • Out-of-distribution questions get unpredictable answers