GhostAI_LiquidSFT v2 (full fine-tune)
On-device Solana wallet assistant — a full-weight fine-tune of LFM2.5-1.2B for
mobile inference (llama.cpp / llama.rn). v2 improves on the v1 LoRA model with a larger,
teacher-augmented + cleaned dataset.
What's new vs v1
- Full-weight fine-tune (8-GPU DDP) instead of LoRA → eval_loss 0.1534 (v1 LoRA: 0.1736)
- Dataset grown to ~78k cleaned rows via grounded augmentation (Qwen3.6 teacher + Google-grounded Solana facts), with: tool-error recovery, multi-step chains, clarification on high-stakes asks, follow-ups, hard negatives, and Ghost AI identity.
- Every tool-call validated against the 172-tool schema; tool args grounded in context (no hallucinated addresses).
Held-out evaluation
| metric | score |
|---|
| Tool name correct | 97.9% |
| Tool full call (name + all args exact) | 85.3% |
| Negatives (no over-trigger) | 88.9% |
| eval_loss | 0.1534 |
Files
| file | quant | size | use |
|---|
GhostAI_LiquidSFT_v2.Q4_0.gguf | Q4_0 | ~664 MB | Phones (ARM) — fastest TTFT+tok/s |
GhostAI_LiquidSFT_v2.Q4_K_M.gguf | Q4_K_M | ~698 MB | desktop balance |
GhostAI_LiquidSFT_v2.Q5_K_M.gguf | Q5_K_M | ~805 MB | higher quality |
GhostAI_LiquidSFT_v2.Q6_K.gguf | Q6_K | ~919 MB | near-lossless |
GhostAI_LiquidSFT_v2.BF16.gguf | BF16 | ~2.2 GB | reference |
⚠️ Serving note (important)
This model is trained train==serve with the on-device tool-catalog system prompt.
Always send that catalog as the system message — with an ad-hoc system prompt, tool-calling
degrades. Tool calls use Hermes format: <tool_call>{"name":...,"arguments":{...}}</tool_call>.
Training
LFM2.5-1.2B-Instruct base · full fine-tune · lr 1e-5 · 2 epochs · eff-batch 256 · bf16 ·
completion_only_loss (user/tool turns masked) · seq 2048 (0% truncation).
