Tool-calling: 1.2B base vs ToolACE LoRA
I fine-tuned LFM2.5-1.2B-Instruct with LoRA on ToolACE to make it better at calling tools with the right arguments. Pick an example to see what each model actually produced.
62.8% → 72.8%exact arguments, 500 held-out ToolACE
+10.0 pts, 95% CI [+6.6, +13.6]
+10.0 pts, 95% CI [+6.6, +13.6]
87.8% → 90.4%BFCL alternative-aware (local metric)
small, tentative gain
small, tentative gain
These are real, recorded outputs from both models on held-out test examples (never seen in training), generated on CPU in float16 with greedy decoding. The published results used float32, so a rerun could occasionally differ. Four examples show the fine-tune fixing the base, one shows both correct, and one shows a fine-tune miss. Metrics are local, not the official BFCL leaderboard, and come from one training seed.