fugu-lite / configs /train_rl.yaml
tahsinsoyak's picture
Upload private Fugu-Lite V1 snapshot
88e15cd verified
Raw
History Blame Contribute Delete
401 Bytes
model:
base_model: Qwen/Qwen3-0.6B
max_length: 512
pooling: last_token
dropout: 0.05
freeze_backbone: true
dtype: auto
training:
seed: 43
epochs: 3
batch_size: 8
gradient_accumulation_steps: 1
learning_rate: 0.0002
weight_decay: 0.0
max_grad_norm: 1.0
estimator: reinforce
rollouts_per_prompt: 8
entropy_coefficient: 0.01
normalize_advantage: false
log_every: 10