3a00ff4d33
- Complete Fastify gateway with 8-stage pipeline - Circuit breaker (opossum) per model tier - Rate limiting per caller - Ban list validation (EN/DE/auto-detected) - TIP validator (SFF-8024, part numbers, wavelengths) - Prometheus metrics - pg-boss async queue - PostgreSQL audit log + review queue - 9 prompt templates (TIP, LinkedIn, ShieldX) - Learning engine scaffolding - Auto-learning: ban-list, few-shot, routing, prompt optimizer
50 lines
1.3 KiB
YAML
50 lines
1.3 KiB
YAML
database_url: "postgresql://llm:llm_secure_password@localhost:5432/llm_gateway"
|
|
gateway_url: "http://localhost:3100"
|
|
ollama_url: "http://192.168.178.169:11434"
|
|
|
|
models:
|
|
qwen_14b_hf: "Qwen/Qwen2.5-14B-Instruct" # HuggingFace model ID — used for general fine-tuning
|
|
qwen_7b_hf: "Qwen/Qwen2.5-7B-Instruct" # For task-specific runs (smaller, faster)
|
|
|
|
training:
|
|
device: "mps" # Apple Silicon MPS — fallback to "cpu" if MPS unavailable
|
|
max_seq_length: 2048
|
|
lora_r: 16
|
|
lora_alpha: 32
|
|
lora_dropout: 0.05
|
|
target_modules:
|
|
- "q_proj"
|
|
- "k_proj"
|
|
- "v_proj"
|
|
- "o_proj"
|
|
- "gate_proj"
|
|
- "up_proj"
|
|
- "down_proj"
|
|
|
|
sft:
|
|
num_epochs: 3
|
|
batch_size: 1
|
|
gradient_accumulation: 8
|
|
learning_rate: 2.0e-4
|
|
warmup_ratio: 0.1
|
|
|
|
dpo:
|
|
num_epochs: 1
|
|
batch_size: 1
|
|
gradient_accumulation: 4
|
|
beta: 0.1 # DPO temperature — higher = more conservative
|
|
learning_rate: 5.0e-5
|
|
|
|
evaluation:
|
|
min_improvement_to_deploy: 0.3 # confidence delta required before deployment
|
|
n_eval_samples: 20
|
|
|
|
output:
|
|
adapters_dir: "adapters"
|
|
models_dir: "models"
|
|
|
|
llama_cpp:
|
|
convert_script: "/opt/homebrew/lib/python3.12/site-packages/llama_cpp/convert_hf_to_gguf.py"
|
|
quantize_binary: "/opt/homebrew/bin/llama-quantize"
|
|
default_quantization: "Q5_K_M"
|