BoyBarley_X

Fine-tuned from BoyBarley-v32 (494M params) via Knowledge Distillation on 50,000 samples.

Training

Config	Value
Base model	BoyBarley/BoyBarley-v32
Dataset	50,000 samples
Categories	SQL, Logic, Math, Reasoning
Epochs	3
Learning rate	1e-5
Precision	BF16
Best val loss	0.0672

Benchmark vs v32

Task	v32	BoyBarley_X
SQL basic	✅	✅ More consistent
SQL aggregate	⚠️	⚠️ Partial
Math basic	⚠️	✅ Improved
Probability	❌	✅ Correct
Syllogism	⚠️	✅ Improved
Trick questions	⚠️	✅ Improved
Fibonacci/Pattern	❌	❌ Still weak
Multi-step math	❌	❌ Still weak

Next: BoyBarley_X2

• Chain-of-thought dataset for math and logic
• Fibonacci and sequence training data
• Compound percentage problems
• Target: 100k samples