chore: remove stale recipes/validator-condition-fft-v3.yaml
Browse files
recipes/validator-condition-fft-v3.yaml
DELETED
|
@@ -1,42 +0,0 @@
|
|
| 1 |
-
# SFT Qwen3-0.6B Validator-Condition (v_c) on section-specific data.
|
| 2 |
-
# Per paper §Combined Validator, v_c critiques only the WHERE/HAVING/CASE conditions.
|
| 3 |
-
model_name_or_path: /home/datht/huggingface/Qwen/Qwen3-0.6B
|
| 4 |
-
model_revision: main
|
| 5 |
-
torch_dtype: bfloat16
|
| 6 |
-
use_flash_attention_2: false
|
| 7 |
-
response_template: "<|im_start|>assistant"
|
| 8 |
-
|
| 9 |
-
dataset_mixer:
|
| 10 |
-
../data/multi-agents/fixed/sft-validator-condition-v3: 1.0
|
| 11 |
-
dataset_splits:
|
| 12 |
-
- train
|
| 13 |
-
- test
|
| 14 |
-
preprocessing_num_workers: 24
|
| 15 |
-
chat_template: "{{'<|im_start|>user\n' + messages['prompt'] + '<|im_end|>\n'}}{{'<|im_start|>assistant\n' + messages['completion'] + '<|im_end|>\n'}}"
|
| 16 |
-
|
| 17 |
-
bf16: true
|
| 18 |
-
do_eval: false
|
| 19 |
-
eval_strategy: "no"
|
| 20 |
-
gradient_accumulation_steps: 8
|
| 21 |
-
gradient_checkpointing: true
|
| 22 |
-
learning_rate: 2.0e-05
|
| 23 |
-
log_level: info
|
| 24 |
-
logging_steps: 10
|
| 25 |
-
logging_strategy: steps
|
| 26 |
-
lr_scheduler_type: cosine
|
| 27 |
-
max_seq_length: 5120
|
| 28 |
-
truncation_side: left
|
| 29 |
-
max_steps: -1
|
| 30 |
-
num_train_epochs: 2
|
| 31 |
-
optim: adamw_torch
|
| 32 |
-
output_dir: output/qwen3-0.6b-bird-validator-condition-sft-v3
|
| 33 |
-
overwrite_output_dir: true
|
| 34 |
-
per_device_train_batch_size: 4
|
| 35 |
-
push_to_hub: false
|
| 36 |
-
remove_unused_columns: true
|
| 37 |
-
report_to:
|
| 38 |
-
- tensorboard
|
| 39 |
-
save_strategy: "epoch"
|
| 40 |
-
save_total_limit: 1
|
| 41 |
-
seed: 42
|
| 42 |
-
tf32: true
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|