GhostAI_LiquidSFT

On-device Solana wallet assistant fine-tuned from LFM2.5-1.2B (Liquid AI) for mobile inference via llama.cpp / llama.rn. Trained to call wallet tools in the Hermes <tool_call>{...}</tool_call> format and to answer Solana / DeFi / wallet-security questions.

Files

File	Quant	Size	Use
GhostAI_LiquidSFT.Q4_K_M.gguf	Q4_K_M	~698 MB	Recommended for phones — best size/quality balance
GhostAI_LiquidSFT.Q5_K_M.gguf	Q5_K_M	~805 MB	Higher quality, modest size bump
GhostAI_LiquidSFT.Q6_K.gguf	Q6_K	~919 MB	Near-lossless vs the 16-bit model
GhostAI_LiquidSFT.BF16.gguf	BF16	~2.2 GB	Full-precision reference for re-quantization

Prompt format (ChatML)

<|im_start|>system
{system prompt with tool catalog}<|im_end|>
<|im_start|>user
{user message}<|im_end|>
<|im_start|>assistant

Tool calls are emitted as:

<tool_call>{"name": "get_sol_balance", "arguments": {"address": "..."}}</tool_call>

The ChatML chat template (with tool + <|im_end|> handling) is embedded in the GGUF.

Training

• Base: unsloth/LFM2.5-1.2B-Instruct
• Method: LoRA SFT (Unsloth), r=32 / α=64, rsLoRA, NEFTune (α=5), bf16
• LoRA targets: q,k,v,out_proj (attention) + in_proj (conv) + w1,w2,w3 (MLP) — all 16 layers
• Data: merged tool-calling dataset (Hermes, 172 tools) + Solana/DeFi/security knowledge base
• Seq len: 2048 (0% truncation) · response-only masking (user and tool turns masked)
• Result: best eval_loss 0.1736, converged at ~1 epoch (early-stopped)

Run with llama.cpp

llama-cli -m GhostAI_LiquidSFT.Q4_K_M.gguf -p "<|im_start|>user\nWhat is my SOL balance?<|im_end|>\n<|im_start|>assistant\n"