Add ORPO example and e2e test (#1572)
Browse files* add example for mistral orpo
* sample_packing: false for orpo
* go to load_dataset (since load_rl_datasets require a transfom_fn, which only dpo uses currently)
- .gitignore +1 -0
- docs/rlhf.qmd +1 -1
- examples/mistral/mistral-qlora-orpo.yml +82 -0
- tests/e2e/test_dpo.py +47 -0
.gitignore
CHANGED
|
@@ -133,6 +133,7 @@ venv/
|
|
| 133 |
ENV/
|
| 134 |
env.bak/
|
| 135 |
venv.bak/
|
|
|
|
| 136 |
|
| 137 |
# Spyder project settings
|
| 138 |
.spyderproject
|
|
|
|
| 133 |
ENV/
|
| 134 |
env.bak/
|
| 135 |
venv.bak/
|
| 136 |
+
venv3.10/
|
| 137 |
|
| 138 |
# Spyder project settings
|
| 139 |
.spyderproject
|
docs/rlhf.qmd
CHANGED
|
@@ -49,7 +49,7 @@ remove_unused_columns: false
|
|
| 49 |
chat_template: chatml
|
| 50 |
datasets:
|
| 51 |
- path: argilla/ultrafeedback-binarized-preferences-cleaned
|
| 52 |
-
type:
|
| 53 |
```
|
| 54 |
|
| 55 |
#### Using local dataset files
|
|
|
|
| 49 |
chat_template: chatml
|
| 50 |
datasets:
|
| 51 |
- path: argilla/ultrafeedback-binarized-preferences-cleaned
|
| 52 |
+
type: chat_template.argilla
|
| 53 |
```
|
| 54 |
|
| 55 |
#### Using local dataset files
|
examples/mistral/mistral-qlora-orpo.yml
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
base_model: mistralai/Mistral-7B-v0.1
|
| 2 |
+
model_type: MistralForCausalLM
|
| 3 |
+
tokenizer_type: LlamaTokenizer
|
| 4 |
+
|
| 5 |
+
load_in_8bit: false
|
| 6 |
+
load_in_4bit: true
|
| 7 |
+
strict: false
|
| 8 |
+
|
| 9 |
+
rl: orpo
|
| 10 |
+
orpo_alpha: 0.1
|
| 11 |
+
remove_unused_columns: false
|
| 12 |
+
|
| 13 |
+
chat_template: chatml
|
| 14 |
+
datasets:
|
| 15 |
+
- path: argilla/ultrafeedback-binarized-preferences-cleaned
|
| 16 |
+
type: chat_template.argilla
|
| 17 |
+
dataset_prepared_path: last_run_prepared
|
| 18 |
+
val_set_size: 0.1
|
| 19 |
+
output_dir: ./mistral-qlora-orpo-out
|
| 20 |
+
|
| 21 |
+
adapter: qlora
|
| 22 |
+
lora_model_dir:
|
| 23 |
+
|
| 24 |
+
sequence_len: 4096
|
| 25 |
+
sample_packing: false
|
| 26 |
+
pad_to_sequence_len: true
|
| 27 |
+
|
| 28 |
+
lora_r: 32
|
| 29 |
+
lora_alpha: 16
|
| 30 |
+
lora_dropout: 0.05
|
| 31 |
+
lora_target_linear: true
|
| 32 |
+
lora_fan_in_fan_out:
|
| 33 |
+
lora_target_modules:
|
| 34 |
+
- gate_proj
|
| 35 |
+
- down_proj
|
| 36 |
+
- up_proj
|
| 37 |
+
- q_proj
|
| 38 |
+
- v_proj
|
| 39 |
+
- k_proj
|
| 40 |
+
- o_proj
|
| 41 |
+
|
| 42 |
+
wandb_project:
|
| 43 |
+
wandb_entity:
|
| 44 |
+
wandb_watch:
|
| 45 |
+
wandb_name:
|
| 46 |
+
wandb_log_model:
|
| 47 |
+
|
| 48 |
+
gradient_accumulation_steps: 4
|
| 49 |
+
micro_batch_size: 2
|
| 50 |
+
num_epochs: 1
|
| 51 |
+
optimizer: adamw_bnb_8bit
|
| 52 |
+
lr_scheduler: cosine
|
| 53 |
+
learning_rate: 0.0002
|
| 54 |
+
|
| 55 |
+
train_on_inputs: false
|
| 56 |
+
group_by_length: false
|
| 57 |
+
bf16: auto
|
| 58 |
+
fp16:
|
| 59 |
+
tf32: false
|
| 60 |
+
|
| 61 |
+
gradient_checkpointing: true
|
| 62 |
+
early_stopping_patience:
|
| 63 |
+
resume_from_checkpoint:
|
| 64 |
+
local_rank:
|
| 65 |
+
logging_steps: 1
|
| 66 |
+
xformers_attention:
|
| 67 |
+
flash_attention: true
|
| 68 |
+
|
| 69 |
+
loss_watchdog_threshold: 5.0
|
| 70 |
+
loss_watchdog_patience: 3
|
| 71 |
+
|
| 72 |
+
warmup_steps: 10
|
| 73 |
+
evals_per_epoch: 4
|
| 74 |
+
eval_table_size:
|
| 75 |
+
eval_max_new_tokens: 128
|
| 76 |
+
saves_per_epoch: 1
|
| 77 |
+
debug:
|
| 78 |
+
deepspeed:
|
| 79 |
+
weight_decay: 0.0
|
| 80 |
+
fsdp:
|
| 81 |
+
fsdp_config:
|
| 82 |
+
special_tokens:
|
tests/e2e/test_dpo.py
CHANGED
|
@@ -158,3 +158,50 @@ class TestDPOLlamaLora(unittest.TestCase):
|
|
| 158 |
|
| 159 |
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
|
| 160 |
assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 158 |
|
| 159 |
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
|
| 160 |
assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
|
| 161 |
+
|
| 162 |
+
@with_temp_dir
|
| 163 |
+
def test_orpo_lora(self, temp_dir):
|
| 164 |
+
# pylint: disable=duplicate-code
|
| 165 |
+
cfg = DictDefault(
|
| 166 |
+
{
|
| 167 |
+
"base_model": "JackFram/llama-68m",
|
| 168 |
+
"tokenizer_type": "LlamaTokenizer",
|
| 169 |
+
"sequence_len": 1024,
|
| 170 |
+
"load_in_8bit": True,
|
| 171 |
+
"adapter": "lora",
|
| 172 |
+
"lora_r": 64,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.1,
|
| 175 |
+
"lora_target_linear": True,
|
| 176 |
+
"special_tokens": {},
|
| 177 |
+
"rl": "orpo",
|
| 178 |
+
"orpo_alpha": 0.1,
|
| 179 |
+
"remove_unused_columns": False,
|
| 180 |
+
"chat_template": "chatml",
|
| 181 |
+
"datasets": [
|
| 182 |
+
{
|
| 183 |
+
"path": "argilla/ultrafeedback-binarized-preferences-cleaned",
|
| 184 |
+
"type": "chat_template.argilla",
|
| 185 |
+
"split": "train",
|
| 186 |
+
},
|
| 187 |
+
],
|
| 188 |
+
"num_epochs": 1,
|
| 189 |
+
"micro_batch_size": 4,
|
| 190 |
+
"gradient_accumulation_steps": 1,
|
| 191 |
+
"output_dir": temp_dir,
|
| 192 |
+
"learning_rate": 0.00001,
|
| 193 |
+
"optimizer": "paged_adamw_8bit",
|
| 194 |
+
"lr_scheduler": "cosine",
|
| 195 |
+
"max_steps": 20,
|
| 196 |
+
"save_steps": 10,
|
| 197 |
+
"warmup_steps": 5,
|
| 198 |
+
"gradient_checkpointing": True,
|
| 199 |
+
"gradient_checkpointing_kwargs": {"use_reentrant": True},
|
| 200 |
+
}
|
| 201 |
+
)
|
| 202 |
+
normalize_config(cfg)
|
| 203 |
+
cli_args = TrainerCliArgs()
|
| 204 |
+
dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
|
| 205 |
+
|
| 206 |
+
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
|
| 207 |
+
assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
|