Strict Exact-Match Accuracy Before and After PEFT LoRA / QLoRA Fine-Tuning
Comparison of unchanged base models versus PEFT LoRA and QLoRA specialist adapters on unseen Tessera MCP tool-call holdout splits (n=150 per split): Llama 3.2 1B rises from 12.0% to 50.7%, Llama 3.2 3B rises from 52.0% to 72.7%, Muse Glimmer 30B frozen test rises from 87.3% to 98.7%, and Muse Glimmer 30B pre-registered fresh holdout rises from 93.3% to 99.3%.
Typed Tool-Call Exact Match on Unseen Holdout Splits (n = 150 each)
Synthetic Tessera MCP specialist task -- proves continuous Task/Task-Object loop mechanics, not general capability
Unchanged Base Model
Promoted PEFT LoRA / QLoRA Adapter
0%
25%
50%
75%
100%
Llama 3.2 1B Instruct
bf16 LoRA r=16 (1x RTX 3060)
12.0%
50.7% (+38.7 pts, R3)
Llama 3.2 3B Instruct
bf16 LoRA r=16 (1x RTX 3060)
52.0%
72.7% (+20.7 pts, R1)
Muse Glimmer 30B (Frozen)
4-bit QLoRA r=16 (free 2x T4)
87.3%
98.7%
Muse Glimmer 30B (Fresh)
Pre-registered SHA-256 holdout
93.3%
99.3%
Guard proof: Llama 3.2 3B Round 2 hit 79.3% exact match but dropped schema validity by 0.025 (limit 0.020), so the pre-registered gate kept Round 1.