How krafton-ai/whale works
WHALE reproduces alternating model-weight and harness optimization experiments across three domains, including Lichess chess puzzles. In each cycle, online rejection-sampling fine-tuning updates the model under the current harness, then Meta-Harness searches for a better harness under the updated model. The chess tree separates puzzle data, environment and reward code, domain harness/runner code, and benchmark/search code. The README documents the optimization loop, but the sampled excerpts do not show the chess modules’ internal calls or their full runtime wiring; those relationships are therefore not inferred.
Main entry points: run WHALE Alternation · meta_harness_chess_puzzle.py Meta-Harness · ray_trainer.py RSFT Training Framework · chess_puzzle_benchmark.py Puzzle Benchmark
Components 15
WHALE Cycle
- WHALE Alternation
domains/chess_puzzles/run - Model Weights
- Updated Harness
Chess Runtime
- Puzzle Data
domains/chess_puzzles/autoharness_chess_puzzle/grpo_dataset.py - Puzzle Environment
domains/chess_puzzles/environments/chess_puzzle/env.py - Base Harness
domains/chess_puzzles/environments/chess_puzzle/base_harness.py - Puzzle Reward
domains/chess_puzzles/environments/chess_puzzle/grpo_reward.py - Chess Harness
domains/chess_puzzles/autoharness_chess_puzzle/harness.py - Chess Runner
domains/chess_puzzles/autoharness_chess_puzzle/runner.py
Harness Search
- Puzzle Benchmark
domains/chess_puzzles/meta_harness/chess_puzzle_benchmark.py - Meta-Harness
domains/chess_puzzles/meta_harness/meta_harness_chess_puzzle.py - Claude Wrapper
domains/chess_puzzles/meta_harness/claude_wrapper.py
Weight Training
- GRPO Dataset
domains/chess_puzzles/autoharness_chess_puzzle/grpo_dataset.py - RSFT Training Framework
domains/chess_puzzles/verl/trainer/ppo/ray_trainer.py
Other components
- Experiment Runner
Connections 8
- Experiment Runner → WHALE Alternation: runs condition
README.md - WHALE Alternation → RSFT Training Framework: trains weights
README.md - RSFT Training Framework → Model Weights: updates weights
README.md - WHALE Alternation → Meta-Harness: searches harness
README.md - Meta-Harness → Updated Harness: improves harness
README.md - WHALE Alternation → Model Weights: alternates updates
README.md - WHALE Alternation → Updated Harness: alternates searches
README.md - Puzzle Benchmark → Updated Harness: evaluates candidates