Sync eval changes in OLMo/ladder-1xC to here

.github/workflows/main.yml

@@ -23,7 +23,7 @@ jobs:
   checks:
     name: ${{ matrix.task.name }}
     runs-on: [ubuntu-latest]
-    timeout-minutes: 5
+    timeout-minutes: 15
     strategy:
       fail-fast: false
       matrix:

src/olmo_eval/metrics.py

@@ -19,12 +19,16 @@ def __init__(self, metric_type="acc") -> None:
         self.metric_type = metric_type
 
         self.add_state("loglikelihoods", default=[], dist_reduce_fx=None)
+        self.add_state("celosses", default=[], dist_reduce_fx=None)
+        self.add_state("bpbs", default=[], dist_reduce_fx=None)
         self.add_state("labels", default=[], dist_reduce_fx=None)
 
     def reset(
         self,
     ):
         self.loglikelihoods = []
+        self.celosses = []
+        self.bpbs = []
         self.labels = []
 
     def update(self, batch: Dict[str, Any], lm_logits: torch.Tensor, dc_lm_logits=None):
@@ -46,6 +50,8 @@ def update(self, batch: Dict[str, Any], lm_logits: torch.Tensor, dc_lm_logits=No
             ]
 
             log_likelihood: torch.Tensor
+            celoss: torch.Tensor
+            bpb: torch.Tensor
             if self.metric_type == "pmi_dc":
                 assert dc_lm_logits is not None
                 # get domain conditional continuation logits: [cont_len, vocab]
@@ -58,19 +64,30 @@ def update(self, batch: Dict[str, Any], lm_logits: torch.Tensor, dc_lm_logits=No
                     torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
                     / torch.gather(dc_lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
                 )
+                celoss = -log_likelihood
+                bpb = -log_likelihood  # the normalization factors cancel out
             elif self.metric_type == "acc" or self.metric_type == "f1":
                 # gather log-probs at continuation token indices
                 log_likelihood = torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-            elif self.metric_type == "len_norm" or self.metric_type == "ce_loss":
+                celoss = (
+                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
+                    / batch["cont_str_len"][idx]
+                )
+                bpb = (
+                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
+                    / batch["cont_byte_len"][idx]
+                    * LOG_2_OF_E
+                )
+            elif self.metric_type in ["len_norm", "ce_loss", "bpb"]:
                 log_likelihood = (
                     torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
                     / batch["cont_str_len"][idx]
                 )
-                if self.metric_type == "ce_loss":
-                    log_likelihood = -log_likelihood
-            elif self.metric_type == "bpb":
-                # bits per byte
-                log_likelihood = (
+                celoss = (
+                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
+                    / batch["cont_str_len"][idx]
+                )
+                bpb = (
                     -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
                     / batch["cont_byte_len"][idx]
                     * LOG_2_OF_E
@@ -84,16 +101,26 @@ def update(self, batch: Dict[str, Any], lm_logits: torch.Tensor, dc_lm_logits=No
                     batch["continuation"][idx].device
                 )
             )
+            self.celosses.append(
+                torch.Tensor((doc_id, cont_id, celoss)).to(batch["continuation"][idx].device)
+            )
+            self.bpbs.append(
+                torch.Tensor((doc_id, cont_id, bpb)).to(batch["continuation"][idx].device)
+            )
             self.labels.append(
                 torch.LongTensor((doc_id, cont_id, batch["label_id"][idx])).to(
                     batch["label_id"][idx].device
                 )
             )
 
-    def compute(self) -> torch.Tensor:
+    def compute(self) -> Dict[str, torch.Tensor]:
+        # Task "suffix" -> tensor
+
         # states should have been synced from all accelerators at this point
         # account for duplicates here because of DistributedSampler compensating for drop_last=False
         loglikelihood_dict: Dict[int, Dict[int, float]] = {}
+        celoss_dict: Dict[int, Dict[int, float]] = {}
+        bpb_dict: Dict[int, Dict[int, float]] = {}
         label_dict = {}
 
         # collect labels
@@ -109,8 +136,29 @@ def compute(self) -> torch.Tensor:
             if int(cont_id.item()) not in loglikelihood_dict[int(doc_id.item())]:
                 loglikelihood_dict[int(doc_id.item())][int(cont_id.item())] = loglikelihood
 
+        # collect celosses
+        for doc_id, cont_id, celoss in self.celosses:
+            if int(doc_id.item()) not in celoss_dict:
+                celoss_dict[int(doc_id.item())] = {}
+
+            if int(cont_id.item()) not in celoss_dict[int(doc_id.item())]:
+                celoss_dict[int(doc_id.item())][int(cont_id.item())] = celoss
+
+        # collect bpbs
+        for doc_id, cont_id, bpb in self.bpbs:
+            if int(doc_id.item()) not in bpb_dict:
+                bpb_dict[int(doc_id.item())] = {}
+
+            if int(cont_id.item()) not in bpb_dict[int(doc_id.item())]:
+                bpb_dict[int(doc_id.item())][int(cont_id.item())] = bpb
+
         # compute acc
         correct = []
+        loglikelihood = []
+        celoss = []
+        bpb = []
+        soft_score = []
+        soft_log_score = []
         preds: Optional[List[float]] = None
         labels: Optional[List[int]] = None
         if self.metric_type == "f1":
@@ -121,37 +169,57 @@ def compute(self) -> torch.Tensor:
             # each doc_id might have a different number of continuation
             num_continuations = len(loglikelihood_dict[doc_id].keys())
             loglikelihoods = torch.tensor([-float("inf")] * num_continuations)
+            celosses = torch.tensor([float("inf")] * num_continuations)
+            bpbs = torch.tensor([float("inf")] * num_continuations)
 
             skip_document = False
             for cont_id in loglikelihood_dict[doc_id]:
                 try:
                     loglikelihoods[cont_id] = loglikelihood_dict[doc_id][cont_id]
+                    celosses[cont_id] = celoss_dict[doc_id][cont_id]
+                    bpbs[cont_id] = bpb_dict[doc_id][cont_id]
                 except IndexError:
                     # We didn't process all of the continuations, so skip this document.
                     skip_document = True
                     break
 
             if skip_document:
                 continue
-            if self.metric_type in ["ce_loss", "bpb"]:
-                correct.append(loglikelihoods[0])  # Only one answer is scored
-            else:
-                correct.append(
-                    1.0 if torch.argmax(loglikelihoods).item() == label_dict[doc_id] else 0.0
-                )
-
-            if self.metric_type == "f1":
+            if self.metric_type == "ce_loss":
+                celoss.append(celosses[0])  # Only one answer is scored
+            elif self.metric_type == "bpb":
+                bpb.append(bpbs[0])  # Only one answer is scored
+            elif self.metric_type == "f1":
                 assert preds is not None
                 assert labels is not None
                 preds.append(torch.argmax(loglikelihoods).item())
                 labels.append(label_dict[doc_id])
+            else:
+                correct.append(
+                    1.0 if torch.argmax(loglikelihoods).item() == label_dict[doc_id] else 0.0
+                )
+                celoss.append(celosses[label_dict[doc_id]].item())
+                bpb.append(bpbs[label_dict[doc_id]].item())
+                soft_score.append(torch.softmax(loglikelihoods, dim=0)[label_dict[doc_id]].item())
+                soft_log_score.append(
+                    torch.log_softmax(loglikelihoods, dim=0)[label_dict[doc_id]].item()
+                )
 
         if self.metric_type == "f1":
             assert preds is not None
             assert labels is not None
             # for NLI tasks, continuations are yes, no, neither, so idx=0 assigned to pos label
             score = f1_score(labels, preds, pos_label=0)
+            return {"f1": torch.tensor(score)}
+        elif self.metric_type == "ce_loss":
+            return {"ce_loss": torch.tensor(sum(celoss) / len(celoss))}
+        elif self.metric_type == "bpb":
+            return {"bpb": torch.tensor(sum(bpb) / len(bpb))}
         else:
-            score = sum(correct) / len(correct)
-
-        return torch.tensor(score)
+            return {
+                self.metric_type: torch.tensor(sum(correct) / len(correct)),
+                "ce_loss": torch.tensor(sum(celoss) / len(celoss)),
+                "bpb": torch.tensor(sum(bpb) / len(bpb)),
+                "soft": torch.tensor(sum(soft_score) / len(soft_score)),
+                "soft_log": torch.tensor(sum(soft_log_score) / len(soft_log_score)),
+            }

src/olmo_eval/oe_eval_tasks/arc_challenge/test_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_challenge:mc", "task_hash": "11abfade7ecce501f3e3e72c937e19cc", "task_config": {"task_name": "arc_challenge:mc", "task_core": "arc_challenge", "limit": 1000000, "split": "test", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Challenge", "dataset_path": "ai2_arc", "dataset_name": "ARC-Challenge", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "arc_challenge:mc::olmes"}}, "current_date": "2024-11-18 22:05:58 UTC", "num_instances": 1172}

src/olmo_eval/oe_eval_tasks/arc_challenge/test_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_challenge", "task_hash": "b122d520ab0cf70114350ecf00c5c811", "task_config": {"task_name": "arc_challenge", "task_core": "arc_challenge", "limit": 1000000, "split": "test", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Challenge", "dataset_path": "ai2_arc", "dataset_name": "ARC-Challenge", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "arc_challenge:rc::olmes"}}, "current_date": "2024-11-18 21:50:18 UTC", "num_instances": 1172}

src/olmo_eval/oe_eval_tasks/arc_challenge/train_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_challenge:mc", "task_hash": "cf2769a2dc6cbea724ff477c3d2543a2", "task_config": {"task_name": "arc_challenge:mc", "task_core": "arc_challenge", "limit": 1000000, "split": "train", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Challenge", "dataset_path": "ai2_arc", "dataset_name": "ARC-Challenge", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "arc_challenge:mc::olmes"}}, "current_date": "2024-11-18 22:05:40 UTC", "num_instances": 1119}

src/olmo_eval/oe_eval_tasks/arc_challenge/train_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_challenge", "task_hash": "9045ed0bd68a7e9ff34cf51ff24828bf", "task_config": {"task_name": "arc_challenge", "task_core": "arc_challenge", "limit": 1000000, "split": "train", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Challenge", "dataset_path": "ai2_arc", "dataset_name": "ARC-Challenge", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "arc_challenge:rc::olmes"}}, "current_date": "2024-11-18 22:05:31 UTC", "num_instances": 1119}

src/olmo_eval/oe_eval_tasks/arc_challenge/val_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_challenge:mc", "task_hash": "a673d7761ce3fc3d5061d72f76755971", "task_config": {"task_name": "arc_challenge:mc", "task_core": "arc_challenge", "limit": 1000000, "split": "validation", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Challenge", "dataset_path": "ai2_arc", "dataset_name": "ARC-Challenge", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "arc_challenge:mc::olmes"}}, "current_date": "2024-11-18 22:05:49 UTC", "num_instances": 299}

src/olmo_eval/oe_eval_tasks/arc_challenge/val_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_challenge", "task_hash": "bd181c90c43b3ef799af2f300ea09cf1", "task_config": {"task_name": "arc_challenge", "task_core": "arc_challenge", "limit": 1000000, "split": "validation", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Challenge", "dataset_path": "ai2_arc", "dataset_name": "ARC-Challenge", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "arc_challenge:rc::olmes"}}, "current_date": "2024-11-18 21:45:07 UTC", "num_instances": 299}

src/olmo_eval/oe_eval_tasks/arc_easy/test_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_easy:mc", "task_hash": "64250ca6fdf0f02e07b539e8efc04922", "task_config": {"task_name": "arc_easy:mc", "task_core": "arc_easy", "limit": 1000000, "split": "test", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Easy", "dataset_path": "ai2_arc", "dataset_name": "ARC-Easy", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"description": "ARC-Easy (MC) using OLMES-v0.1", "regimes": ["OLMES-v0.1"], "alias": "arc_easy:mc::olmes"}}, "current_date": "2024-11-18 22:06:33 UTC", "num_instances": 2376}

src/olmo_eval/oe_eval_tasks/arc_easy/test_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_easy", "task_hash": "ccbbd993c851d3300140d81ffec0e397", "task_config": {"task_name": "arc_easy", "task_core": "arc_easy", "limit": 1000000, "split": "test", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": {"description": null}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Easy", "dataset_path": "ai2_arc", "dataset_name": "ARC-Easy", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"description": "ARC-Easy (RC) using OLMES-v0.1", "regimes": ["OLMES-v0.1"], "alias": "arc_easy:rc::olmes"}}, "current_date": "2024-11-18 21:50:27 UTC", "num_instances": 2376}

src/olmo_eval/oe_eval_tasks/arc_easy/train_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_easy:mc", "task_hash": "afa7e96b485c4e4481b3b9b817faac36", "task_config": {"task_name": "arc_easy:mc", "task_core": "arc_easy", "limit": 1000000, "split": "train", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Easy", "dataset_path": "ai2_arc", "dataset_name": "ARC-Easy", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"description": "ARC-Easy (MC) using OLMES-v0.1", "regimes": ["OLMES-v0.1"], "alias": "arc_easy:mc::olmes"}}, "current_date": "2024-11-18 22:06:16 UTC", "num_instances": 2251}

src/olmo_eval/oe_eval_tasks/arc_easy/train_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_easy", "task_hash": "4a5241b308edb45d7b9eab594093c519", "task_config": {"task_name": "arc_easy", "task_core": "arc_easy", "limit": 1000000, "split": "train", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": {"description": null}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Easy", "dataset_path": "ai2_arc", "dataset_name": "ARC-Easy", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"description": "ARC-Easy (RC) using OLMES-v0.1", "regimes": ["OLMES-v0.1"], "alias": "arc_easy:rc::olmes"}}, "current_date": "2024-11-18 22:06:07 UTC", "num_instances": 2251}

src/olmo_eval/oe_eval_tasks/arc_easy/val_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_easy:mc", "task_hash": "443bd52f752399615d01c853a8d7386c", "task_config": {"task_name": "arc_easy:mc", "task_core": "arc_easy", "limit": 1000000, "split": "validation", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Easy", "dataset_path": "ai2_arc", "dataset_name": "ARC-Easy", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"description": "ARC-Easy (MC) using OLMES-v0.1", "regimes": ["OLMES-v0.1"], "alias": "arc_easy:mc::olmes"}}, "current_date": "2024-11-18 22:06:24 UTC", "num_instances": 570}

src/olmo_eval/oe_eval_tasks/arc_easy/val_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "arc_easy", "task_hash": "0045e4f588a617cbe9ee5a4ae8ca1ce5", "task_config": {"task_name": "arc_easy", "task_core": "arc_easy", "limit": 1000000, "split": "validation", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_per_char", "random_subsample_seed": 1234, "context_kwargs": {"description": null}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": 1000000}, "native_id_field": "id", "fewshot_source": "OLMES:ARC-Easy", "dataset_path": "ai2_arc", "dataset_name": "ARC-Easy", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"description": "ARC-Easy (RC) using OLMES-v0.1", "regimes": ["OLMES-v0.1"], "alias": "arc_easy:rc::olmes"}}, "current_date": "2024-11-18 21:45:15 UTC", "num_instances": 570}

src/olmo_eval/oe_eval_tasks/boolq/train_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "boolq:mc", "task_hash": "a92ca849d7efd331110145eb71e4fc09", "task_config": {"task_name": "boolq:mc", "task_core": "boolq", "limit": 1000000, "split": "train", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": null}, "native_id_field": "idx", "fewshot_source": "OLMES:BoolQ", "dataset_path": "super_glue", "dataset_name": "boolq", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "boolq:mc::olmes"}}, "current_date": "2024-11-18 22:06:52 UTC", "num_instances": 9427}

src/olmo_eval/oe_eval_tasks/boolq/train_rc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "boolq", "task_hash": "ec8729b372d310aaf3a222f37a7af7b9", "task_config": {"task_name": "boolq", "task_core": "boolq", "limit": 1000000, "split": "train", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": null}, "native_id_field": "idx", "fewshot_source": "OLMES:BoolQ", "dataset_path": "super_glue", "dataset_name": "boolq", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "boolq:rc::olmes"}}, "current_date": "2024-11-18 22:06:41 UTC", "num_instances": 9427}

src/olmo_eval/oe_eval_tasks/boolq/val_mc_5shot/config.json

@@ -0,0 +1 @@
+{"task_name": "boolq:mc", "task_hash": "d88f45757f4a8c3802b7274857894a90", "task_config": {"task_name": "boolq:mc", "task_core": "boolq", "limit": 1000000, "split": "validation", "num_shots": 5, "fewshot_seed": 1234, "primary_metric": "acc_raw", "random_subsample_seed": 1234, "context_kwargs": {}, "generation_kwargs": {}, "metric_kwargs": {"uncond_docid_offset": null}, "native_id_field": "idx", "fewshot_source": "OLMES:BoolQ", "dataset_path": "super_glue", "dataset_name": "boolq", "use_chat_format": null, "version": 0, "revision": null, "metadata": {"regimes": ["OLMES-v0.1"], "alias": "boolq:mc::olmes"}}, "current_date": "2024-11-18 22:07:01 UTC", "num_instances": 3270}