[Chat] fix eval: add in training evaluation

applications/ColossalChat/README.md

@@ -529,7 +529,7 @@ Coati is developed by ColossalAI Team:
 - [Fazzie](https://fazzie-key.cool/about/index.html) Contributing to the algorithm and development for SFT.
 - [ofey404](https://github.com/ofey404) Contributing to both front-end and back-end development.
 - [Wenhao Chen](https://github.com/CWHer) Contributing to subsequent code enhancements and performance improvements.
-- [Anbang Ye](https://github.com/YeAnbang) Contributing to the refactored PPO version with updated acceleration framework. Add support for DPO, SimPO.
+- [Anbang Ye](https://github.com/YeAnbang) Contributing to the refactored PPO version with updated acceleration framework. Add support for DPO, SimPO, ORPO.
 
 The PhD student from [(HPC-AI) Lab](https://ai.comp.nus.edu.sg/) also contributed a lot to this project.
 - [Zangwei Zheng](https://github.com/zhengzangw)
@@ -579,6 +579,36 @@ We also appreciate the valuable suggestions provided by [Jian Hu](https://github
   journal = {GitHub repository},
   howpublished = {\url{https://github.com/XueFuzhao/InstructionWild}},
 }
+
+@misc{meng2024simposimplepreferenceoptimization,
+      title={SimPO: Simple Preference Optimization with a Reference-Free Reward},
+      author={Yu Meng and Mengzhou Xia and Danqi Chen},
+      year={2024},
+      eprint={2405.14734},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2405.14734},
+}
+
+@misc{rafailov2023directpreferenceoptimizationlanguage,
+      title={Direct Preference Optimization: Your Language Model is Secretly a Reward Model},
+      author={Rafael Rafailov and Archit Sharma and Eric Mitchell and Stefano Ermon and Christopher D. Manning and Chelsea Finn},
+      year={2023},
+      eprint={2305.18290},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2305.18290},
+}
+
+@misc{hong2024orpomonolithicpreferenceoptimization,
+      title={ORPO: Monolithic Preference Optimization without Reference Model},
+      author={Jiwoo Hong and Noah Lee and James Thorne},
+      year={2024},
+      eprint={2403.07691},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2403.07691},
+}
 ```
 
 ## Licenses

applications/ColossalChat/coati/dataset/loader.py

@@ -28,6 +28,8 @@ def load_tokenized_dataset(
     Each instance of dataset is a dictionary with
     `{'input_ids': List[int], 'labels': List[int], sequence: str}` format.
     """
+    if not dataset_paths:
+        return None
     mode_map = kwargs.get("mode_map", {"train": "train", "dev": "validation", "test": "test"})
     assert mode in tuple(mode_map), f"Unsupported mode {mode}, it must be in {tuple(mode_map)}"
 

applications/ColossalChat/coati/trainer/dpo.py

@@ -2,6 +2,7 @@
 Dpo trainer
 """
 
+import os
 from typing import Any, Optional
 
 import torch
@@ -324,7 +325,7 @@ def _eval(self, epoch: int):
                     chosen_loss_mask[:, 1:],
                     reject_loss_mask[:, 1:],
                 )
-                reward_accuracies = (chosen_rewards > rejected_rewards).float()
+                reward_accuracies = (chosen_rewards > rejected_rewards).float().mean()
                 loss = losses.mean()
                 loss_mean = all_reduce_mean(tensor=loss)
                 chosen_rewards_mean = all_reduce_mean(tensor=chosen_rewards)
@@ -343,4 +344,7 @@ def _eval(self, epoch: int):
         for tag in ["loss", "chosen_rewards", "rejected_rewards", "accuracy", "margin"]:
             msg = msg + f"{tag}: {self.accumulative_meter.get(tag)}\n"
         self.coordinator.print_on_master(msg)
+        os.makedirs(self.save_dir, exist_ok=True)
+        with open(os.path.join(self.save_dir, f"eval_result_epoch{epoch}.txt"), "w") as f:
+            f.write(msg)
         step_bar.close()

applications/ColossalChat/coati/trainer/orpo.py

@@ -2,6 +2,7 @@
 Orpo trainer
 """
 
+import os
 from typing import Any, Optional
 
 import torch
@@ -269,11 +270,10 @@ def _eval(self, epoch: int):
                 batch_size = chosen_input_ids.size()[0]
                 actor_out = self.model(
                     input_ids=torch.cat([chosen_input_ids, reject_input_ids]),
-                    labels=torch.cat([chosen_input_ids, reject_input_ids]),
                     attention_mask=torch.cat([chosen_attention_mask, reject_attention_mask]),
                 )
+                torch.autograd.set_detect_anomaly(True)
                 actor_all_logits = actor_out["logits"].to(torch.float32)
-                chosen_nll = torch.mean(actor_out["loss"][:batch_size]).to(dtype=torch.bfloat16)
                 actor_chosen_logits = actor_all_logits[:batch_size]
                 actor_reject_logits = actor_all_logits[batch_size:]
                 logprob_actor_chosen = calc_masked_log_probs(
@@ -283,14 +283,22 @@ def _eval(self, epoch: int):
                 logprob_actor_reject = calc_masked_log_probs(
                     actor_reject_logits, reject_input_ids, reject_loss_mask[:, 1:]
                 )
-
-                odds_ratio_loss, log_odds_ratio = self.odds_ratio_loss_fn(logprob_actor_chosen, logprob_actor_reject)
-
+                chosen_logits = actor_chosen_logits[:, :-1, :].contiguous().view(-1, actor_chosen_logits.size(-1))
+                label_chosen = chosen_input_ids[:, 1:].contiguous()
+                label_chosen_masked = (
+                    label_chosen.masked_fill(chosen_loss_mask[:, 1:] == 0, -100).view(-1).contiguous().detach()
+                )
+                # label_chosen[chosen_loss_mask[:, 1:] == 0] = -100
+                chosen_nll = self.sft_loss_fn(chosen_logits, label_chosen_masked).to(dtype=torch.bfloat16)
+                odds_ratio_loss, log_odds_ratio = self.odds_ratio_loss_fn(
+                    logprob_actor_chosen, logprob_actor_reject, chosen_loss_mask[:, 1:], reject_loss_mask[:, 1:]
+                )
                 loss = chosen_nll - odds_ratio_loss * self.lam
+                step_bar.set_description(f"Epoch {epoch + 1}/{self.max_epochs} Loss: {loss.detach().cpu().item():.4f}")
 
-                chosen_rewards = torch.mean(logprob_actor_chosen).item()
-                rejected_rewards = torch.mean(logprob_actor_reject).item()
-                reward_accuracies = (log_odds_ratio > 0).float().mean().item()
+                chosen_rewards = torch.sum(logprob_actor_chosen) / torch.sum(chosen_loss_mask[:, 1:])
+                rejected_rewards = torch.sum(logprob_actor_reject) / torch.sum(reject_loss_mask[:, 1:])
+                reward_accuracies = torch.sum((log_odds_ratio > 0).float()) / torch.sum(log_odds_ratio != 0)
 
                 # sync
                 loss_mean = all_reduce_mean(tensor=loss)
@@ -303,37 +311,11 @@ def _eval(self, epoch: int):
                 self.accumulative_meter.add("log_odds_ratio", log_odds_ratio.to(torch.float16).mean().item())
                 self.accumulative_meter.add("accuracy", reward_accuracies_mean.to(torch.float16).item())
 
-                # logging
-                if self.writer and is_rank_0():
-                    self.writer.add_scalar("eval/loss", self.accumulative_meter.get("loss"), self.num_train_step)
-                    self.writer.add_scalar("train/lr", self.optimizer.param_groups[0]["lr"], self.num_train_step)
-                    self.writer.add_scalar(
-                        "train/chosen_rewards", self.accumulative_meter.get("chosen_rewards"), self.num_train_step
-                    )
-                    self.writer.add_scalar(
-                        "train/rejected_rewards",
-                        self.accumulative_meter.get("rejected_rewards"),
-                        self.num_train_step,
-                    )
-                    self.writer.add_scalar(
-                        "train/log",
-                        self.accumulative_meter.get("chosen_rewards") - self.accumulative_meter.get("rejected_rewards"),
-                        self.num_train_step,
-                    )
-                    self.writer.add_scalar(
-                        "train/accuracy",
-                        self.accumulative_meter.get("accuracy"),
-                        self.num_train_step,
-                    )
-                    self.writer.add_scalar(
-                        "train/log_odds_ratio",
-                        self.accumulative_meter.get("log_odds_ratio"),
-                        self.num_train_step,
-                    )
-                    self.step_bar.update()
-
         msg = "Evaluation Result:\n"
         for tag in ["loss", "chosen_rewards", "rejected_rewards", "log_odds_ratio", "accuracy"]:
             msg = msg + f"{tag}: {self.accumulative_meter.get(tag)}\n"
         self.coordinator.print_on_master(msg)
+        os.makedirs(self.save_dir, exist_ok=True)
+        with open(os.path.join(self.save_dir, f"eval_result_epoch{epoch}.txt"), "w") as f:
+            f.write(msg)
         step_bar.close()

applications/ColossalChat/coati/trainer/rm.py

@@ -237,6 +237,7 @@ def _eval(self, epoch):
                 + f"distance: {self.accumulative_meter.get('chosen_rewards')-self.accumulative_meter.get('rejected_rewards')}\n"
             )
             self.coordinator.print_on_master(msg)
+            os.makedirs(self.save_dir, exist_ok=True)
             with open(os.path.join(self.save_dir, f"eval_result_epoch{epoch}.txt"), "w") as f:
                 f.write(msg)
             step_bar.close()

applications/ColossalChat/coati/trainer/sft.py

@@ -167,6 +167,7 @@ def _eval(self, epoch: int):
             for tag in ["loss"]:
                 msg = msg + f"{tag}: {self.accumulative_meter.get(tag)}\n"
             self.coordinator.print_on_master(msg)
+            os.makedirs(self.save_dir, exist_ok=True)
             with open(os.path.join(self.save_dir, f"eval_result_epoch{epoch}.txt"), "w") as f:
                 f.write(msg)
             step_bar.close()

applications/ColossalChat/examples/training_scripts/train_dpo.py

@@ -176,6 +176,19 @@ def train(args):
         collate_fn=data_collator,
         distributed_sampler_cls=StatefulDistributedSampler,
     )
+    eval_dataloader = None
+    if args.eval_dataset:
+        eval_dataset = load_tokenized_dataset(dataset_paths=args.eval_dataset, mode="dev")
+        eval_data_collator = DataCollatorForPreferenceDataset(tokenizer=tokenizer, max_length=args.max_length)
+
+        eval_dataloader = plugin.prepare_dataloader(
+            dataset=eval_dataset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            drop_last=True,
+            collate_fn=eval_data_collator,
+            distributed_sampler_cls=StatefulDistributedSampler,
+        )
 
     num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps
     if args.warmup_steps is None:
@@ -260,7 +273,7 @@ def train(args):
 
     trainer.fit(
         train_preference_dataloader=train_dataloader,
-        eval_preference_dataloader=None,
+        eval_preference_dataloader=eval_dataloader,
         log_dir=args.log_dir,
         use_wandb=args.use_wandb,
     )
@@ -309,6 +322,7 @@ def train(args):
     parser.add_argument("--model_type", type=str, default=None)
     parser.add_argument("--tokenizer_dir", type=str, default=None)
     parser.add_argument("--dataset", nargs="+", default=[])
+    parser.add_argument("--eval_dataset", nargs="+", default=[])
     parser.add_argument(
         "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
     )

applications/ColossalChat/examples/training_scripts/train_orpo.py

@@ -164,6 +164,19 @@ def train(args):
         distributed_sampler_cls=StatefulDistributedSampler,
     )
 
+    eval_dataloader = None
+    if args.eval_dataset:
+        eval_dataset = load_tokenized_dataset(dataset_paths=args.eval_dataset, mode="dev")
+        eval_data_collator = DataCollatorForPreferenceDataset(tokenizer=tokenizer, max_length=args.max_length)
+        eval_dataloader = plugin.prepare_dataloader(
+            dataset=eval_dataset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            drop_last=True,
+            collate_fn=eval_data_collator,
+            distributed_sampler_cls=StatefulDistributedSampler,
+        )
+
     num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps
     if args.warmup_steps is None:
         args.warmup_steps = int(args.max_epochs * 0.025 * (len(train_dataloader) // args.accumulation_steps))
@@ -242,7 +255,7 @@ def train(args):
 
     trainer.fit(
         train_preference_dataloader=train_dataloader,
-        eval_preference_dataloader=None,
+        eval_preference_dataloader=eval_dataloader,
         log_dir=args.log_dir,
         use_wandb=args.use_wandb,
     )
@@ -288,6 +301,7 @@ def train(args):
     parser.add_argument("--model_type", type=str, default=None)
     parser.add_argument("--tokenizer_dir", type=str, default=None)
     parser.add_argument("--dataset", nargs="+", default=[])
+    parser.add_argument("--eval_dataset", nargs="+", default=[])
     parser.add_argument(
         "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
     )

applications/ColossalChat/examples/training_scripts/train_rm.py

@@ -173,6 +173,20 @@ def train(args):
         collate_fn=data_collator,
         distributed_sampler_cls=StatefulDistributedSampler,
     )
+
+    eval_dataloader = None
+    if args.eval_dataset:
+        eval_dataset = load_tokenized_dataset(dataset_paths=args.eval_dataset, mode="dev")
+        eval_data_collator = DataCollatorForPreferenceDataset(tokenizer=tokenizer, max_length=args.max_length)
+        eval_dataloader = plugin.prepare_dataloader(
+            dataset=eval_dataset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            drop_last=True,
+            collate_fn=eval_data_collator,
+            distributed_sampler_cls=StatefulDistributedSampler,
+        )
+
     num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps
     math.ceil(args.max_epochs * num_update_steps_per_epoch)
 
@@ -297,6 +311,7 @@ def train(args):
     parser.add_argument("--pretrain", type=str, default=None)
     parser.add_argument("--tokenizer_dir", type=str, default=None)
     parser.add_argument("--dataset", nargs="+", default=[])
+    parser.add_argument("--eval_dataset", nargs="+", default=[])
     parser.add_argument(
         "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
     )

applications/ColossalChat/examples/training_scripts/train_sft.py

@@ -173,6 +173,21 @@ def train(args):
         collate_fn=data_collator,
         distributed_sampler_cls=StatefulDistributedSampler,
     )
+
+    eval_dataloader = None
+    if args.eval_dataset:
+        eval_dataset = load_tokenized_dataset(dataset_paths=args.eval_dataset, mode="dev")
+        eval_data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer, max_length=args.max_len)
+
+        eval_dataloader = plugin.prepare_dataloader(
+            dataset=eval_dataset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            drop_last=True,
+            collate_fn=eval_data_collator,
+            distributed_sampler_cls=StatefulDistributedSampler,
+        )
+
     coordinator.print_on_master(
         f"Max CUDA memory after data loader: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB"
     )
@@ -255,7 +270,7 @@ def train(args):
 
     trainer.fit(
         train_dataloader=train_dataloader,
-        eval_dataloader=None,
+        eval_dataloader=eval_dataloader,
         log_dir=args.log_dir,
         use_wandb=args.use_wandb,
     )
@@ -300,6 +315,7 @@ def train(args):
     parser.add_argument("--pretrain", type=str, default=None)
     parser.add_argument("--tokenizer_dir", type=str, default=None)
     parser.add_argument("--dataset", nargs="+", default=[])
+    parser.add_argument("--eval_dataset", nargs="+", default=[])
     parser.add_argument(
         "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
     )