support ORPO

Former-commit-id: f44a4c27e2461cdaa1b16865f597a31033c0e6d9

README.md

@@ -68,16 +68,18 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 ## Changelog
 
+[24/03/31] We supported **[ORPO](https://arxiv.org/abs/2403.07691)**. See `examples/lora_single_gpu` for usage.
+
 [24/03/21] Our paper "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" is available at arXiv!
 
 [24/03/20] We supported **FSDP+QLoRA** that fine-tunes a 70B model on 2x24GB GPUs. See `examples/fsdp_qlora` for usage.
 
+<details><summary>Full Changelog</summary>
+
 [24/03/13] We supported **[LoRA+](https://arxiv.org/abs/2402.12354)**. See `examples/extras/loraplus` for usage.
 
 [24/03/07] We supported gradient low-rank projection (**[GaLore](https://arxiv.org/abs/2403.03507)**) algorithm. See `examples/extras/galore` for usage.
 
-<details><summary>Full Changelog</summary>
-
 [24/03/07] We integrated **[vLLM](https://github.com/vllm-project/vllm)** for faster and concurrent inference. Try `--infer_backend vllm` to enjoy **270%** inference speed. (LoRA is not yet supported, merge it first.)
 
 [24/02/28] We supported weight-decomposed LoRA (**[DoRA](https://arxiv.org/abs/2402.09353)**). Try `--use_dora` to activate DoRA training.
@@ -165,6 +167,7 @@ You also can add a custom chat template to [template.py](src/llmtuner/data/templ
 | Reward Modeling        | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | PPO Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | DPO Training           | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
+| ORPO Training          | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
 > [!NOTE]
 > Use `--quantization_bit 4` argument to enable QLoRA.

README_zh.md

@@ -68,16 +68,18 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 ## 更新日志
 
+[24/03/31] 我们支持了 **[ORPO](https://arxiv.org/abs/2403.07691)**。详细用法请参照 `examples/lora_single_gpu`。
+
 [24/03/21] 我们的论文 "[LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://arxiv.org/abs/2403.13372)" 可在 arXiv 上查看！
 
 [24/03/20] 我们支持了能在 2x24GB GPU 上微调 70B 模型的 **FSDP+QLoRA**。详细用法请参照 `examples/fsdp_qlora`。
 
+<details><summary>展开日志</summary>
+
 [24/03/13] 我们支持了 **[LoRA+](https://arxiv.org/abs/2402.12354)**。详细用法请参照 `examples/extras/loraplus`。
 
 [24/03/07] 我们支持了梯度低秩投影（**[GaLore](https://arxiv.org/abs/2403.03507)**）算法。详细用法请参照 `examples/extras/galore`。
 
-<details><summary>展开日志</summary>
-
 [24/03/07] 我们集成了 **[vLLM](https://github.com/vllm-project/vllm)** 以实现极速并发推理。请使用 `--infer_backend vllm` 来获得 **270%** 的推理速度。（尚不支持 LoRA，请先合并权重。）
 
 [24/02/28] 我们支持了 **[DoRA](https://arxiv.org/abs/2402.09353)** 微调。请使用 `--use_dora` 参数进行 DoRA 微调。
@@ -165,6 +167,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 | 奖励模型训练            | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | PPO 训练               | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 | DPO 训练               | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
+| ORPO 训练              | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: |
 
 > [!NOTE]
 > 请使用 `--quantization_bit 4` 参数来启用 QLoRA 训练。

data/README.md

@@ -34,6 +34,8 @@ If you are using a custom dataset, please provide your dataset definition in the
 
 Given above, you can use the custom dataset via specifying `--dataset dataset_name`.
 
+----
+
 Currently we support dataset in **alpaca** or **sharegpt** format, the dataset in alpaca format should follow the below format:
 
 ```json
@@ -84,6 +86,10 @@ For the preference datasets, the `response` column should be a string list whose
 }
 ```
 
+Remember to set `"ranking": true` for the preference datasets.
+
+----
+
 The dataset in sharegpt format should follow the below format:
 
 ```json

data/README_zh.md

@@ -34,6 +34,8 @@
 
 添加后可通过指定 `--dataset 数据集名称` 参数使用自定义数据集。
 
+----
+
 该项目目前支持两种格式的数据集：**alpaca** 和 **sharegpt**，其中 alpaca 格式的数据集按照以下方式组织：
 
 ```json
@@ -84,6 +86,10 @@
 }
 ```
 
+添加偏好数据集需要额外指定 `"ranking": true`。
+
+----
+
 而 sharegpt 格式的数据集按照以下方式组织：
 
 ```json

examples/lora_single_gpu/README.md

@@ -1,8 +1,9 @@
 Usage:
 
 - `pretrain.sh`: do pre-train (optional)
-- `sft.sh`: do supervised fine-tune
+- `sft.sh`: do supervised fine-tuning
 - `reward.sh`: do reward modeling (must after sft.sh)
 - `ppo.sh`: do PPO training (must after sft.sh and reward.sh)
 - `dpo.sh`: do DPO training (must after sft.sh)
+- `orpo.sh`: do ORPO training
 - `predict.sh`: do predict (must after sft.sh and dpo.sh)

examples/lora_single_gpu/orpo.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+
+CUDA_VISIBLE_DEVICES=0 python ../../src/train_bash.py \
+    --stage orpo \
+    --do_train \
+    --model_name_or_path meta-llama/Llama-2-7b-hf \
+    --dataset comparison_gpt4_en \
+    --dataset_dir ../../data \
+    --template default \
+    --finetuning_type lora \
+    --lora_target q_proj,v_proj \
+    --output_dir ../../saves/LLaMA2-7B/lora/orpo \
+    --overwrite_cache \
+    --overwrite_output_dir \
+    --cutoff_len 1024 \
+    --preprocessing_num_workers 16 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --logging_steps 10 \
+    --warmup_steps 20 \
+    --save_steps 100 \
+    --eval_steps 100 \
+    --evaluation_strategy steps \
+    --load_best_model_at_end \
+    --learning_rate 1e-5 \
+    --num_train_epochs 1.0 \
+    --max_samples 1000 \
+    --val_size 0.1 \
+    --plot_loss \
+    --fp16

src/llmtuner/data/__init__.py

@@ -1,6 +1,15 @@
+from .collator import PairwiseDataCollatorWithPadding
 from .loader import get_dataset
 from .template import Template, get_template_and_fix_tokenizer, templates
 from .utils import Role, split_dataset
 
 
-__all__ = ["get_dataset", "Template", "get_template_and_fix_tokenizer", "templates", "Role", "split_dataset"]
+__all__ = [
+    "PairwiseDataCollatorWithPadding",
+    "get_dataset",
+    "Template",
+    "get_template_and_fix_tokenizer",
+    "templates",
+    "Role",
+    "split_dataset",
+]

src/llmtuner/data/collator.py

@@ -0,0 +1,51 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Sequence, Tuple
+
+import torch
+from transformers import DataCollatorForSeq2Seq
+
+
+@dataclass
+class PairwiseDataCollatorWithPadding(DataCollatorForSeq2Seq):
+    r"""
+    Data collator for pairwise data.
+    """
+
+    def _pad_labels(self, batch: torch.Tensor, positions: List[Tuple[int, int]]) -> torch.Tensor:
+        r"""
+        Masks out the input ids except for the responses.
+        """
+        padded_labels = []
+        for feature, (prompt_len, answer_len) in zip(batch, positions):
+            if self.tokenizer.padding_side == "left":
+                start, end = feature.size(0) - answer_len, feature.size(0)
+            else:
+                start, end = prompt_len, prompt_len + answer_len
+            padded_tensor = self.label_pad_token_id * torch.ones_like(feature)
+            padded_tensor[start:end] = feature[start:end]
+            padded_labels.append(padded_tensor)
+        return torch.stack(padded_labels, dim=0).contiguous()  # in contiguous memory
+
+    def __call__(self, features: Sequence[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
+        r"""
+        Pads batched data to the longest sequence in the batch.
+
+        We generate 2 * n examples where the first n examples represent chosen examples and
+        the last n examples represent rejected examples.
+        """
+        concatenated_features = []
+        label_positions = []
+        for key in ("chosen_ids", "rejected_ids"):
+            for feature in features:
+                prompt_len, answer_len = len(feature["prompt_ids"]), len(feature[key])
+                concatenated_features.append(
+                    {
+                        "input_ids": feature["prompt_ids"] + feature[key],
+                        "attention_mask": [1] * (prompt_len + answer_len),
+                    }
+                )
+                label_positions.append((prompt_len, answer_len))
+
+        batch = super().__call__(concatenated_features)
+        batch["labels"] = self._pad_labels(batch["input_ids"], label_positions)
+        return batch

src/llmtuner/data/loader.py

@@ -117,7 +117,6 @@ def get_dataset(
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo"],
-    # split: Optional[str] = "train", # TODO: add split
 ) -> Union["Dataset", "IterableDataset"]:
     template = get_template_and_fix_tokenizer(tokenizer, data_args.template)
     if data_args.train_on_prompt and template.efficient_eos:
@@ -138,6 +137,9 @@ def get_dataset(
     with training_args.main_process_first(desc="load dataset"):
         all_datasets = []
         for dataset_attr in get_dataset_list(data_args):
+            if (stage == "rm" and dataset_attr.ranking is False) or (stage != "rm" and dataset_attr.ranking is True):
+                raise ValueError("The dataset is not applicable in the current training stage.")
+
             all_datasets.append(load_single_dataset(dataset_attr, model_args, data_args))
         dataset = merge_dataset(all_datasets, data_args, training_args)
 

src/llmtuner/data/preprocess.py

@@ -23,23 +23,25 @@ def preprocess_pretrain_dataset(
 ) -> Dict[str, List[List[int]]]:
     # build grouped texts with format `X1 X2 X3 ...` if packing is enabled
     text_examples = [messages[0]["content"] + tokenizer.eos_token for messages in examples["prompt"]]
-    if not data_args.packing:
-        return tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len)
 
-    tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
-    concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
-    total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
-    block_size = data_args.cutoff_len
-    # we drop the small remainder, and if the total_length < block_size, we exclude this batch
-    total_length = (total_length // block_size) * block_size
-    # split by chunks of cutoff_len
-    result = {
-        k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
-        for k, t in concatenated_examples.items()
-    }
-    if data_args.template == "gemma":
-        for i in range(len(result["input_ids"])):
-            result["input_ids"][i][0] = tokenizer.bos_token_id
+    if not data_args.packing:
+        if data_args.template == "gemma":
+            text_examples = [tokenizer.bos_token + example for example in text_examples]
+
+        result = tokenizer(text_examples, add_special_tokens=False, max_length=data_args.cutoff_len)
+    else:
+        tokenized_examples = tokenizer(text_examples, add_special_tokens=False)
+        concatenated_examples = {k: list(chain(*tokenized_examples[k])) for k in tokenized_examples.keys()}
+        total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
+        block_size = data_args.cutoff_len
+        total_length = (total_length // block_size) * block_size
+        result = {
+            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        if data_args.template == "gemma":
+            for i in range(len(result["input_ids"])):
+                result["input_ids"][i][0] = tokenizer.bos_token_id
 
     return result
 

src/llmtuner/data/utils.py

@@ -44,7 +44,7 @@ def checksum(data_files: List[str], file_sha1: Optional[str] = None) -> None:
 def infer_max_len(source_len: int, target_len: int, max_len: int, reserved_label_len: int) -> Tuple[int, int]:
     max_target_len = int(max_len * (target_len / (source_len + target_len)))
     max_target_len = max(max_target_len, reserved_label_len)
-    max_source_len = max_len - max_target_len
+    max_source_len = max_len - min(max_target_len, target_len)
     return max_source_len, max_target_len
 
 

src/llmtuner/extras/callbacks.py

@@ -134,6 +134,7 @@ class LogCallback(TrainerCallback):
             eval_loss=state.log_history[-1].get("eval_loss", None),
             predict_loss=state.log_history[-1].get("predict_loss", None),
             reward=state.log_history[-1].get("reward", None),
+            accuracy=state.log_history[-1].get("rewards/accuracies", None),
             learning_rate=state.log_history[-1].get("learning_rate", None),
             epoch=state.log_history[-1].get("epoch", None),
             percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,

src/llmtuner/extras/constants.py

@@ -39,9 +39,12 @@ TRAINING_STAGES = {
     "Reward Modeling": "rm",
     "PPO": "ppo",
     "DPO": "dpo",
+    "ORPO": "orpo",
     "Pre-Training": "pt",
 }
 
+STAGES_USE_PAIR_DATA = ["rm", "dpo", "orpo"]
+
 V_HEAD_WEIGHTS_NAME = "value_head.bin"
 
 V_HEAD_SAFE_WEIGHTS_NAME = "value_head.safetensors"

src/llmtuner/hparams/finetuning_args.py

@@ -110,6 +110,10 @@ class RLHFArguments:
         default=0.0,
         metadata={"help": "The supervised fine-tuning loss coefficient in DPO training."},
     )
+    orpo_beta: float = field(
+        default=0.1,
+        metadata={"help": "The beta (lambda) parameter in ORPO loss representing the weight of the SFT loss."},
+    )
     ppo_buffer_size: int = field(
         default=1,
         metadata={"help": "The number of mini-batches to make experience buffer in a PPO optimization step."},
@@ -209,7 +213,7 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
         default=False,
         metadata={"help": "Whether or not to train model in purely bf16 precision (without AMP)."},
     )
-    stage: Literal["pt", "sft", "rm", "ppo", "dpo"] = field(
+    stage: Literal["pt", "sft", "rm", "ppo", "dpo", "orpo"] = field(
         default="sft",
         metadata={"help": "Which stage will be performed in training."},
     )

src/llmtuner/train/dpo/trainer.py

@@ -74,7 +74,7 @@ class CustomDPOTrainer(DPOTrainer):
         create_custom_scheduler(self.args, num_training_steps, optimizer)
         return super().create_scheduler(num_training_steps, optimizer)
 
-    def sft_loss(self, chosen_logits: torch.FloatTensor, chosen_labels: torch.LongTensor) -> torch.Tensor:
+    def sft_loss(self, chosen_logits: "torch.FloatTensor", chosen_labels: "torch.LongTensor") -> "torch.Tensor":
         r"""
         Computes supervised cross-entropy loss of given labels under the given logits.
 
@@ -85,8 +85,8 @@ class CustomDPOTrainer(DPOTrainer):
         return -all_logps
 
     def concatenated_forward(
-        self, model: "PreTrainedModel", batch: Dict[str, torch.Tensor]
-    ) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
         batch_copied = BatchEncoding({k: v.detach().clone() for k, v in batch.items()})  # avoid error
 
         all_logits = model(
@@ -107,9 +107,9 @@ class CustomDPOTrainer(DPOTrainer):
     def get_batch_loss_metrics(
         self,
         model: "PreTrainedModel",
-        batch: Dict[str, torch.Tensor],
+        batch: Dict[str, "torch.Tensor"],
         train_eval: Literal["train", "eval"] = "train",
-    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
+    ) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
         r"""
         Computes the DPO loss and other metrics for the given batch of inputs for train or test.
         """
@@ -142,21 +142,22 @@ class CustomDPOTrainer(DPOTrainer):
             reference_chosen_logps,
             reference_rejected_logps,
         )
+        batch_loss = losses.mean()
         if self.ftx_gamma > 1e-6:
             batch_size = batch["input_ids"].size(0) // 2
             chosen_labels, _ = batch["labels"].split(batch_size, dim=0)
-            losses += self.ftx_gamma * self.sft_loss(policy_chosen_logits, chosen_labels)
+            batch_loss += self.ftx_gamma * self.sft_loss(policy_chosen_logits, chosen_labels).mean()
 
         reward_accuracies = (chosen_rewards > rejected_rewards).float()
 
         prefix = "eval_" if train_eval == "eval" else ""
-        metrics[f"{prefix}rewards/chosen"] = chosen_rewards.cpu().mean()
-        metrics[f"{prefix}rewards/rejected"] = rejected_rewards.cpu().mean()
-        metrics[f"{prefix}rewards/accuracies"] = reward_accuracies.cpu().mean()
-        metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).cpu().mean()
-        metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.detach().cpu().mean()
-        metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.detach().cpu().mean()
-        metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.detach().cpu().mean()
-        metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.detach().cpu().mean()
+        metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.cpu().mean()
+        metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.cpu().mean()
+        metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.cpu().mean()
+        metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).cpu().mean()
+        metrics["{}logps/rejected".format(prefix)] = policy_rejected_logps.detach().cpu().mean()
+        metrics["{}logps/chosen".format(prefix)] = policy_chosen_logps.detach().cpu().mean()
+        metrics["{}logits/rejected".format(prefix)] = policy_rejected_logits.detach().cpu().mean()
+        metrics["{}logits/chosen".format(prefix)] = policy_chosen_logits.detach().cpu().mean()
 
-        return losses.mean(), metrics
+        return batch_loss, metrics

src/llmtuner/train/dpo/workflow.py

@@ -2,13 +2,12 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import get_dataset, split_dataset
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
 from ...extras.constants import IGNORE_INDEX
 from ...extras.ploting import plot_loss
 from ...hparams import ModelArguments
 from ...model import load_model, load_tokenizer
 from ..utils import create_modelcard_and_push, create_ref_model
-from .collator import DPODataCollatorWithPadding
 from .trainer import CustomDPOTrainer
 
 
@@ -29,7 +28,7 @@ def run_dpo(
     dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
 
-    data_collator = DPODataCollatorWithPadding(
+    data_collator = PairwiseDataCollatorWithPadding(
         tokenizer=tokenizer,
         pad_to_multiple_of=8,
         label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
@@ -64,7 +63,7 @@ def run_dpo(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "accuracy"])
 
     # Evaluation
     if training_args.do_eval:

src/llmtuner/train/orpo/__init__.py

@@ -0,0 +1,4 @@
+from .workflow import run_orpo
+
+
+__all__ = ["run_orpo"]

src/llmtuner/train/orpo/trainer.py

@@ -0,0 +1,150 @@
+from collections import defaultdict
+from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import Trainer
+from trl import DPOTrainer
+from trl.trainer.utils import disable_dropout_in_model
+
+from ...extras.constants import IGNORE_INDEX
+from ..utils import create_custom_optimzer, create_custom_scheduler
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+    from ...hparams import FinetuningArguments
+
+
+class CustomORPOTrainer(DPOTrainer):
+    def __init__(
+        self,
+        model: Union["PreTrainedModel", "torch.nn.Module"],
+        finetuning_args: "FinetuningArguments",
+        disable_dropout: bool = True,
+        **kwargs,
+    ):
+        if disable_dropout:
+            disable_dropout_in_model(model)
+
+        self.finetuning_args = finetuning_args
+        self.reference_free = False
+        self.use_dpo_data_collator = True  # hack to avoid warning
+        self.generate_during_eval = False  # disable at evaluation
+        self.label_pad_token_id = IGNORE_INDEX
+        self.padding_value = 0
+        self.is_encoder_decoder = model.config.is_encoder_decoder
+        self.precompute_ref_log_probs = False
+        self._precomputed_train_ref_log_probs = False
+        self._precomputed_eval_ref_log_probs = False
+        self._peft_has_been_casted_to_bf16 = False
+
+        self.beta = finetuning_args.orpo_beta
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+
+        Trainer.__init__(self, model=model, **kwargs)
+
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
+    ) -> "torch.optim.lr_scheduler.LRScheduler":
+        create_custom_scheduler(self.args, num_training_steps, optimizer)
+        return super().create_scheduler(num_training_steps, optimizer)
+
+    def sft_loss(self, chosen_logits: "torch.FloatTensor", chosen_labels: "torch.LongTensor") -> "torch.Tensor":
+        r"""
+        Computes supervised cross-entropy loss of given labels under the given logits.
+
+        Returns:
+            A tensor of shape (batch_size,) containing the cross-entropy loss of each samples.
+        """
+        all_logps = self.get_batch_logps(chosen_logits, chosen_labels, average_log_prob=True)
+        return -all_logps
+
+    # Borrowed from:
+    # https://github.com/huggingface/trl/blob/0ee349dcd43b0f4b3169449f16751c38ac4a609f/trl/trainer/orpo_trainer.py#L592
+    def odds_ratio_loss(
+        self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor"
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes ORPO's odds ratio (OR) loss.
+
+        Args:
+            policy_chosen_logps: Log probabilities of the policy model for the chosen responses. Shape: (batch_size,)
+            policy_rejected_logps: Log probabilities of the policy model for the rejected responses. Shape: (batch_size,)
+
+        Returns:
+            A tuple of five tensors: (losses, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen).
+        """
+
+        # Derived from Eqs. (4) and (7) from https://arxiv.org/abs/2403.07691 by using log identities and exp(log(P(y|x)) = P(y|x)
+        log_odds = (chosen_logps - rejected_logps) - (
+            torch.log(1 - torch.exp(chosen_logps)) - torch.log(1 - torch.exp(rejected_logps))
+        )
+        ratio = F.logsigmoid(log_odds)
+        losses = self.beta * ratio
+
+        chosen_rewards = self.beta * chosen_logps.detach()
+        rejected_rewards = self.beta * rejected_logps.detach()
+
+        return losses, chosen_rewards, rejected_rewards, ratio, log_odds
+
+    def concatenated_forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        all_logits = model(
+            input_ids=batch["input_ids"], attention_mask=batch["attention_mask"], return_dict=True
+        ).logits.to(torch.float32)
+
+        all_logps = self.get_batch_logps(
+            all_logits,
+            batch["labels"],
+            average_log_prob=False,
+            label_pad_token_id=self.label_pad_token_id,
+        )
+        batch_size = batch["input_ids"].size(0) // 2
+        chosen_logps, rejected_logps = all_logps.split(batch_size, dim=0)
+        chosen_logits, rejected_logits = all_logits.split(batch_size, dim=0)
+        return chosen_logps, rejected_logps, chosen_logits, rejected_logits
+
+    def get_batch_loss_metrics(
+        self,
+        model: "PreTrainedModel",
+        batch: Dict[str, "torch.Tensor"],
+        train_eval: Literal["train", "eval"] = "train",
+    ) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
+        r"""
+        Computes the ORPO loss and other metrics for the given batch of inputs for train or test.
+        """
+        metrics = {}
+        chosen_logps, rejected_logps, chosen_logits, rejected_logits = self.concatenated_forward(model, batch)
+
+        losses, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen = self.odds_ratio_loss(
+            chosen_logps, rejected_logps
+        )
+        batch_size = batch["input_ids"].size(0) // 2
+        chosen_labels, _ = batch["labels"].split(batch_size, dim=0)
+        sft_loss = self.sft_loss(chosen_logits, chosen_labels)
+        batch_loss = (sft_loss - losses).mean()
+
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.cpu().mean()
+        metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.cpu().mean()
+        metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.cpu().mean()
+        metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).cpu().mean()
+        metrics["{}logps/rejected".format(prefix)] = rejected_logps.detach().cpu().mean()
+        metrics["{}logps/chosen".format(prefix)] = chosen_logps.detach().cpu().mean()
+        metrics["{}logits/rejected".format(prefix)] = rejected_logits.detach().cpu().mean()
+        metrics["{}logits/chosen".format(prefix)] = chosen_logits.detach().cpu().mean()
+        metrics["{}sft_loss".format(prefix)] = sft_loss.detach().cpu().mean()
+        metrics["{}log_odds_ratio".format(prefix)] = log_odds_ratio.detach().cpu().mean()
+        metrics["{}log_odds_chosen".format(prefix)] = log_odds_chosen.detach().cpu().mean()
+
+        return batch_loss, metrics

src/llmtuner/train/orpo/workflow.py

@@ -0,0 +1,68 @@
+# Inspired by: https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama_2/scripts/dpo_llama2.py
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
+from ...extras.constants import IGNORE_INDEX
+from ...extras.ploting import plot_loss
+from ...hparams import ModelArguments
+from ...model import load_model, load_tokenizer
+from ..utils import create_modelcard_and_push
+from .trainer import CustomORPOTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments
+
+
+def run_orpo(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer = load_tokenizer(model_args)
+    dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+
+    data_collator = PairwiseDataCollatorWithPadding(
+        tokenizer=tokenizer,
+        pad_to_multiple_of=8,
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
+    )
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for pairwise dataset
+
+    # Initialize our Trainer
+    trainer = CustomORPOTrainer(
+        model=model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **split_dataset(dataset, data_args, training_args),
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "accuracy"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Create model card
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

src/llmtuner/train/rm/workflow.py

@@ -2,13 +2,12 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import get_dataset, split_dataset
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
 from ...extras.callbacks import FixValueHeadModelCallback
 from ...extras.misc import fix_valuehead_checkpoint
 from ...extras.ploting import plot_loss
 from ...model import load_model, load_tokenizer
 from ..utils import create_modelcard_and_push
-from .collator import PairwiseDataCollatorWithPadding
 from .metric import compute_accuracy
 from .trainer import PairwiseTrainer
 

src/llmtuner/train/tuner.py

@@ -9,6 +9,7 @@ from ..extras.logging import get_logger
 from ..hparams import get_infer_args, get_train_args
 from ..model import load_model_and_tokenizer
 from .dpo import run_dpo
+from .orpo import run_orpo
 from .ppo import run_ppo
 from .pt import run_pt
 from .rm import run_rm
@@ -36,6 +37,8 @@ def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: Optional[List["Tra
         run_ppo(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
     elif finetuning_args.stage == "dpo":
         run_dpo(model_args, data_args, training_args, finetuning_args, callbacks)
+    elif finetuning_args.stage == "orpo":
+        run_orpo(model_args, data_args, training_args, finetuning_args, callbacks)
     else:
         raise ValueError("Unknown task.")
 

src/llmtuner/webui/common.py

@@ -11,6 +11,7 @@ from ..extras.constants import (
     DEFAULT_MODULE,
     DEFAULT_TEMPLATE,
     PEFT_METHODS,
+    STAGES_USE_PAIR_DATA,
     SUPPORTED_MODELS,
     TRAINING_STAGES,
     DownloadSource,
@@ -127,7 +128,7 @@ def load_dataset_info(dataset_dir: str) -> Dict[str, Dict[str, Any]]:
 
 def list_dataset(dataset_dir: str = None, training_stage: str = list(TRAINING_STAGES.keys())[0]) -> "gr.Dropdown":
     dataset_info = load_dataset_info(dataset_dir if dataset_dir is not None else DEFAULT_DATA_DIR)
-    ranking = TRAINING_STAGES[training_stage] in ["rm", "dpo"]
+    ranking = TRAINING_STAGES[training_stage] in STAGES_USE_PAIR_DATA
     datasets = [k for k, v in dataset_info.items() if v.get("ranking", False) == ranking]
     return gr.Dropdown(value=[], choices=datasets)