immortaltatsu/ghostai-lfm2.5-1.2b-app-v2

GhostAI LFM2.5-1.2B — app-contract (eval-2)

This model is a 1.2 billion parameter, on-device tool-calling model developed by immortaltatsu for the GhostWallet Solana application. It is fine-tuned to emit single Hermes-style tool calls based on the app's system prompt and a retrieved tool catalog slice, then provide a one-line answer from the tool's result. The model is built upon LiquidAI's LFM2.5-1.2B-Thinking base.

Key Capabilities and Performance

• Tool Calling: Designed to emit a single <tool_call> block without prose, followed by a one-line answer after tool execution.
• Enhanced Validation: Achieves 91.1% for arguments passing app validation and 100% prompt-injection resistance, showing significant improvements over previous versions.
• Training Data: Trained on a 1000-question iOS UI evaluation set and a comprehensive 174-tool teacher corpus, with weights from epoch 1 to prevent overfitting.
• On-Device Optimization: Available in Q4_K_M.gguf format (698 MB) for efficient execution on devices using llama.rn or llama.cpp.
• Security: Demonstrates 0 confirm-gate bypasses and no planted content reaching tool arguments in evaluations.

Limitations

• Grounding and Multi-turn: Grounding is 55.9% and multi-turn capabilities are at 25%, indicating areas for potential improvement.
• Retrieval Bottleneck: Tool retrieval remains a limiting factor, with the gold tool offered only about 10.7% of the time in the single-turn set.
• Synthetic Data: Training relies solely on synthetic data, lacking real user transcripts.
• Host App Dependency: Requires the host app to manage tool retrieval, argument validation, and value-moving confirmation gates for safe operation.