35 lines
834 B
YAML
35 lines
834 B
YAML
|
# LoRA adapter configuration — standalone reference for inference/export
|
||
|
|
# Matches the training config in slm_jitna_v0.1.yaml
|
||
|
|
|
||
|
|
peft_type: "LORA"
|
||
|
|
task_type: "CAUSAL_LM"
|
||
|
|
|
||
|
|
# Adapter dimensions
|
||
|
|
r: 16
|
||
|
|
lora_alpha: 32
|
||
|
|
lora_dropout: 0.05
|
||
|
|
bias: "none"
|
||
|
|
use_rslora: true
|
||
|
|
|
||
|
|
# Target modules (all projection layers)
|
||
|
|
target_modules:
|
||
|
|
- q_proj
|
||
|
|
- k_proj
|
||
|
|
- v_proj
|
||
|
|
- o_proj
|
||
|
|
- gate_proj
|
||
|
|
- up_proj
|
||
|
|
- down_proj
|
||
|
|
|
||
|
|
# Inference settings
|
||
|
|
inference_mode: false # set true for inference-only loading
|
||
|
|
|
||
|
|
# Merge settings (for GGUF export)
|
||
|
|
merge_before_export: true
|
||
|
|
export_format: "Q4_K_M" # 4-bit quantized GGUF for Ollama
|
||
|
|
|
||
|
|
# Paths (relative to repo root)
|
||
|
|
adapter_path: "models/adapters/jitna_v0.1"
|
||
|
|
merged_path: "models/merged/jitna_v0.1"
|
||
|
|
gguf_path: "models/gguf/delentia-jitna-v0.1-Q4_K_M.gguf"
|