Merge branch 'fix-seq2seq' of github.com:allenai/RL4LMs into fix-seq2seq

rl4lms/envs/text_generation/registry.py

@@ -19,6 +19,7 @@
     ToTTo,
     WMT14PreprocessedEnDe,
     WMT16NewsOnlyDatasetEnDe,
+    DailyDialog
 )
 from rl4lms.data_pools.text_generation_pool import TextGenPool
 from rl4lms.envs.text_generation.alg_wrappers import wrap_onpolicy_alg
@@ -94,6 +95,7 @@ class DataPoolRegistry:
         "wmt16newsonly": WMT16NewsOnlyDatasetEnDe,
         "iwslt2017en_de": IWSLT2017EnDe,
         "crd3": CRD3DialogueGeneration,
+        "daily_dialog": DailyDialog
     }
 
     @classmethod

scripts/training/task_configs/dialog/gpt2_supervised.yml

@@ -0,0 +1,50 @@
+tokenizer:
+  model_name: gpt2
+  padding_side: left
+  truncation_side: left
+  pad_token_as_eos_token: True
+  max_length: 64
+
+datapool:
+  id: "daily_dialog"
+  args:
+    context_size: 5
+
+alg:
+  id: supervised
+  training_args:
+    per_device_train_batch_size:
+      expand: True
+      values: [32]
+    logging_steps: 200
+    num_train_epochs: 10
+    lr_scheduler_type: "constant"
+    learning_rate: 
+      expand: True 
+      values: [0.00001]
+    save_total_limit: 1
+  model_type: causal
+  model_name: gpt2
+  generation_kwargs:
+    do_sample: True
+    min_length: 10
+    max_new_tokens: 50
+    post_processing_fn: null
+
+train_evaluation:
+  eval_batch_size: 256
+  metrics:
+    - id: diversity
+      args: {}
+    - id: meteor
+      args: {}
+    - id: rouge
+    - id: bleu
+      args: {}
+    - id: bert_score
+      args:
+        language: en
+    - id: sacre_bleu
+      args:
+        tokenize: "intl"
+