bt-sim-slm

benchmark: held-out accuracy per run and training loss

A small causal LM fine-tuned from LiquidAI/LFM2.5-230M-Base (modified: fine-tuned, files changed) to turn a JSON task spec (plus a node catalog, the previous XML and validator errors on a retry) into BehaviorTree.CPP v4 XML.

  • Training data: synthetic (task spec, XML) pairs labelled by a validator; rows longer than 2560 tokens were skipped; 1 epoch.
  • Held-out score against a rule-based generator on synthetic specs: valid 0.975, exact 0.98. This shows the model learned the rules; it is NOT evaluated on real LLM output.
  • This version: v5: also trained on long plans (heavy / max heavy specs); on 20 held-out rows each, heavy 0.90 valid / 0.70 exact, max 0.90 / 0.80 (v4: 0.55 / 0.35, 0.80 / 0.50)
  • License: the base model's LFM Open License v1.0 (copy in LICENSE); attribution to Liquid AI, Inc. retained; commercial use is restricted above the license's revenue threshold.
Downloads last month
49
Safetensors
Model size
0.2B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for larmkaixian/bt-sim-slm

Finetuned
(22)
this model