BENCH-001: Ternary vs f16/bf16 Comparison (Minimal Research)

[gHashTag]

Task

Minimal BENCH-001: Compare ternary, f16, bf16 on small MLP using existing .tri specs.

Goal: First honest "ternary vs binary" validation for TRI-27 algorithms.

Approach

• Use existing specs/algo/mlp.tri (4 → 8 → 3 architecture)
• Compare: ternary {-1,0,+1} vs fp32 vs fp16 vs bf16
• Metrics: accuracy, MSE loss, memory usage

Experiments

Experiment	Input	Weights	Metrics
Baseline	[1,0,0,0]	fp32	loss, accuracy
Ternary	[1,0,0,0]	{-1,0,+1}	loss, accuracy
FP16	[1,0,0,0]	fp16	loss, accuracy
BF16	[1,0,0,0]	bf16	loss, accuracy

Implementation

File: src/bench_ternary_vs_binary.zig

// 1. fp32 baseline (already exists in test_mlp_semantic.zig)
// 2. Quantize weights to ternary {-1,0,+1}
// 3. Quantize weights to fp16/bf16
// 4. Compare outputs

Success Criteria

• Ternary weights achieve <5% accuracy loss vs fp32
• BF16 within 1% of fp32 (expected from BENCH-004)
• Research log with raw numbers (not marketing)

Dependencies

• Uses: specs/algo/mlp.tri, src/formats.zig (GF16 encode/decode)
• Related: BENCH-001: Number format benchmarks #479 (BENCH-001 parent issue)

Timeline

Target: April 5, 2026 (3 days)

Output

results/bench_001_summary.csv:

format,accuracy,loss,memory_bytes
fp32,95.0,0.123,4
ternary,92.0,0.145,1
fp16,94.8,0.124,2
bf16,94.9,0.123,2

[gHashTag]

Translation task appears complete. All Russian text across trinity-w1 repository should be English now.