53 lines
2.1 KiB
Markdown
53 lines
2.1 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: LiquidAI/LFM2.5-1.2B
|
|
tags:
|
|
- gguf
|
|
- llama.cpp
|
|
- lfm2
|
|
- on-device
|
|
- tool-calling
|
|
- solana
|
|
- wallet-assistant
|
|
- full-finetune
|
|
library_name: gguf
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# 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).
|