File size: 1,733 Bytes
49a4262 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | # โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Chinese BabyLM Evaluation Pipeline Config
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Models to evaluate.
# Each entry needs:
# path โ HuggingFace repo ID or local directory path
# backend โ one of: causal, mlm, mntp, enc_dec_mask, enc_dec_prefix
models:
- path: /home/haihu2/scratch/models/Qwen3-0.6B
backend: causal
# - path: /path/to/local/model
# backend: mlm
# Tasks to run. Comment out any group or individual task to skip it.
tasks:
# NLU Track โ zero-shot minimal pairs
zero_shot:
- zhoblimp
- hanzi_structure
- hanzi_pinyin
# Cog Track โ fMRI brain encoding
cogbench:
- word_fmri
- fmri
# Fine-tuning Track โ CLUE tasks
finetune:
- afqmc
- ocnli
- tnews
- cluewsc2020
# Directories
eval_dir: evaluation_data # where prepare_chinese_data.py puts data
results_dir: results # where eval results are written
# Save items containing UNK tokens for hanzi track tasks
save_item_with_unk: true
# Fine-tuning hyperparameters
# Global defaults are applied first; per-task overrides are merged on top.
finetune_hparams:
lr: 3.0e-5
batch_size: 32
max_epochs: 10
sequence_length: 128
seed: 42
task_overrides:
afqmc:
lr: 3.0e-5
batch_size: 32
max_epochs: 10
ocnli:
lr: 3.0e-5
batch_size: 32
max_epochs: 10
tnews:
lr: 3.0e-5
batch_size: 32
max_epochs: 10
cluewsc2020:
lr: 3.0e-5
batch_size: 32
max_epochs: 30
|