config.yaml.example

# LLM-Imitate pipeline configuration
# Edit these values before running each step.

# --- Data paths ---
thread_path: "data/raw/inbox"
output_dir: "data/processed"

# --- Persona target ---
# Must match sender_name exactly as it appears in the Instagram export (after mojibake fix).
target_sender: "Target Name"          # the person to imitate
my_sender: "Your Name"              # your name in the thread

# --- Parsing filters ---
include_shares: false        # include link/share messages as text
include_reactions_only: false

# --- Turn grouping & dataset ---
turn_gap_seconds: 60         # merge rapid-fire msgs from same sender
context_messages: 6          # prior turns included as context
train_split: 0.95
min_completion_chars: 1      # skip empty completions

# --- PII review ---
# Messages matching these patterns are flagged in pii_review.jsonl for manual review.
pii_flag_enabled: true

# --- Model & training ---
# Set max_train_samples to a small number (e.g. 200) for a quick smoke test.
# Set to null or remove to use the full dataset.
max_train_samples: null
base_model: "unsloth/Llama-3.1-8B-Instruct-bnb-4bit"
adapter_output_dir: "models/adapters"
lora_rank: 16
lora_alpha: 32
lora_dropout: 0.05
epochs: 1
batch_size: 2
gradient_accumulation_steps: 4
learning_rate: 2.0e-4
max_seq_length: 2048
warmup_ratio: 0.03
save_steps: 9999
eval_steps: 9999

# --- Inference ---
gguf_output_dir: "models/gguf"
gguf_quant: "Q4_K_M"
ollama_model_name: "persona-imitate"
default_temperature: 0.7
vibe_check_candidates: 3