phila-ft-v5-GGUF

phila-ft-v5 (Q4_K_M)

Gate-only fine-tune of Llama 3.2 3B for phila, a silence-first group chat agent. v5 adds targeted already-corrected scenario augmentation on top of v3's proven 3,799-record dataset.

Training

• Method: QLoRA r=16 a=32, 4-bit quantization
• Examples: 4,780 gate-only (v3 base + 70 targeted already-corrected)
• GPU: Vast.ai RTX 4090
• Base model: meta-llama/Llama-3.2-3B-Instruct

Results

With the restructured prompt (NEVER SPEAK first, 7 worked examples):

Key improvements over v3: "already corrected" scenarios fixed (100% pass rate), false speaks reduced from 15 to 4, holdout accuracy +1.6pp.

Usage

# Create Ollama model
ollama create phila-ft-v5 -f Modelfile

Modelfile:

FROM phila-ft-v5-unsloth.Q4_K_M.gguf
TEMPLATE "{{- if .System }}<|start_header_id|>system<|end_header_id|>

{{ .System }}<|eot_id|>{{ end }}{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>

{{ .Content }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>

"
PARAMETER top_p 0.52
PARAMETER num_predict 64
PARAMETER stop <|start_header_id|>
PARAMETER stop <|end_header_id|>
PARAMETER stop <|eot_id|>
PARAMETER temperature 0.1

Project

github.com/onblueroses/phila