immortaltatsu/ghostai-lfm-app
The immortaltatsu/ghostai-lfm-app is a 1.2 billion parameter LFM2.5-1.2B-Thinking model, fine-tuned by immortaltatsu for direct integration with the GhostWallet mobile app's on-device tool contract. This model excels at generating parseable and valid tool calls, achieving 98.2% accuracy in tool selection and argument construction when the correct tool is available. It is specifically optimized for precise tool invocation within a mobile application environment, demonstrating significant improvements over prior checkpoints in contract adherence and argument validation.
Loading preview...
GhostAI LFM App-Contract SFT
This model, immortaltatsu/ghostai-lfm-app, is a specialized fine-tune of the LFM2.5-1.2B-Thinking model, featuring 1.2 billion parameters and a 32768 token context length. Its primary purpose is to generate tool calls that precisely match the GhostWallet mobile app's on-device tool contract, using the Hermes <tool_call> block format.
Key Capabilities & Improvements
- High Tool Call Parseability & Validity: Achieves 92.9% parseable tool calls and 87.5% argument validity against the app's validator in real-world scenarios.
- Near-Perfect Tool Selection: When the correct tool is made available in the catalog, the model demonstrates 98.2% accuracy in selecting the expected tool and 94.6% valid arguments, effectively solving the problem of tool selection and argument construction.
- Optimized for App Integration: Directly targets the app's contract, resolving issues where previous models emitted incorrect argument formats.
- Reduced Completion Tokens: Generates more concise responses, averaging 40.7 completion tokens per turn compared to 63 in the baseline.
Training Details
The model was trained on 5,407 app-contract traces covering all 174 app tools, using a teacher-distilled approach from Qwen3.6-35B-A3B. It employs assistant-only loss to ensure accurate reproduction of rendered tool-call history.
Known Weaknesses
- Injection Resistance Regression: A slight decrease from 100% to 94.4% in injection resistance, though critical invariants (no confirm-gate bypass, no planted content in tool args) are maintained.
- Weak Multi-turn Performance: Multi-turn interactions, including follow-ups and pronoun resolution, remain unreliable at 16.7%.
Usage
Designed to be served with llama.cpp, the model expects the app's assembled system prompt and outputs tool calls in the specified JSON format. The base model is under the LFM Open License, with commercial use restrictions for revenues above $10M annually.