# PERMANENCE — training configuration.
#
# Four-stage pipeline: supervised warmup -> format-coverage gate ->
# GRPO -> held-out evaluation. Single NVIDIA T4 GPU (16 GB VRAM).
# End-to-end runtime ~1 h 20 min.
#
# See docs/METHODS.md for the rationale behind every hyperparameter
# on this page.

model_name: unsloth/Llama-3.2-3B-Instruct-bnb-4bit

# 300 prompts x group_size=4 rollouts = 1 200 total training episodes.
total_episodes: 300

# Group size chosen so per-device batch equals group size; this avoids
# Unsloth's auto-batching inflating memory on a 16 GB T4.
group_size: 4

# Standard TRL defaults for PPO-style optimisation on LoRA adapters.
learning_rate: 4.0e-5

# KL coefficient against the SFT reference model. The TRL default 0.04
# was chosen deliberately — a looser constraint (0.02 in a pilot) lets
# the policy drift away from its warmup-established calibration once
# the curriculum phases in harder scenarios.
kl_coefficient: 0.04

# Two inner PPO updates per generation batch. Trades a small amount of
# off-policy drift for faster convergence.
ppo_epochs: 2

gradient_clip: 1.0
lora_r: 16
lora_alpha: 16
load_in_4bit: true
max_seq_length: 1088
output_dir: ./training/artifacts
checkpoint_frequency: 150
warmup_sft_epochs: 10
format_reward_cutoff: 300
eval_episodes: 36
eval_seed_offset: 50000

# Domain filter applied to the curriculum sampler. Training focuses on
# the devtools domain (filesystem / git / database). The meridian
# domain is also registered — it demonstrates that the pipeline is
# domain-agnostic — but is not sampled during training.
domain: devtools