19 lines
546 B
YAML
19 lines
546 B
YAML
|
seed: 42
|
||
|
|
output_dir: /workspace/research/tmp/qwen3_3p1b_a0p85b_moe_candidate_v1_recovery_cn_v1
|
||
|
|
model_path: /workspace/research/tmp/qwen3_3p1b_a0p85b_moe_candidate_v1
|
||
|
|
tokenizer_path: /workspace/research/projects/reft-ai/models/Qwen/Qwen3-4B-Instruct-2507
|
||
|
|
dataset_paths:
|
||
|
|
- data/chinese_expanded/recovery_cn_merged_v1.jsonl
|
||
|
|
max_seq_len: 320
|
||
|
|
batch_size: 1
|
||
|
|
gradient_accumulation_steps: 4
|
||
|
|
num_epochs: 4
|
||
|
|
learning_rate: 3.0e-7
|
||
|
|
warmup_ratio: 0.05
|
||
|
|
weight_decay: 0.0
|
||
|
|
max_grad_norm: 1.0
|
||
|
|
save_every: 25
|
||
|
|
log_every: 1
|
||
|
|
bf16: true
|
||
|
|
gradient_checkpointing: false
|