Upload 3 files

.gitattributes

@@ -42,3 +42,4 @@ musique/musique_ans_v1.0_train.jsonl filter=lfs diff=lfs merge=lfs -text
 musique/musique_full_v1.0_dev.jsonl filter=lfs diff=lfs merge=lfs -text
 musique/musique_full_v1.0_test.jsonl filter=lfs diff=lfs merge=lfs -text
 musique/musique_full_v1.0_train.jsonl filter=lfs diff=lfs merge=lfs -text
+qrag-ft-e5-on-hotpotqa/eval_seed42.jsonl filter=lfs diff=lfs merge=lfs -text

qrag-ft-e5-on-hotpotqa/config.yaml

@@ -0,0 +1,178 @@
+envs:
+  max_steps: 2
+  task: hotpotqa
+  data_path: ./datasets/data_sources/hotpotqa
+  sort_by_index: false
+  positions_processor: none
+  positions_processor_dict:
+    none:
+      _target_: envs.text_env.AbsolutePositionProcessor
+    absolute:
+      _target_: envs.text_env.AbsolutePositionProcessor
+    random:
+      _target_: envs.text_env.RandomPositionProcessor
+      max_chunks_count: 2000
+    relative:
+      _target_: envs.text_env.RelativePositionProcessor
+      step_size: 20
+  train_dataset:
+    _target_: envs.RetrievalHotPotQA
+    path: ${envs.data_path}
+    split: train
+    seed: ${seed}
+  test_dataset:
+    _target_: envs.RetrievalHotPotQA
+    path: ${envs.data_path}
+    split: eval
+    seed: ${seed}
+  env:
+    _target_: envs.qa_env.QAEnv
+    max_steps: ${envs.max_steps}
+    action_embed_length: ${max_action_length}
+    separator: ' [SEP] '
+    sort_by_index: ${envs.sort_by_index}
+    positions_processor: ${envs.positions_processor_dict.${envs.positions_processor}}
+    feedback_model:
+      _target_: rl.feedback.GroundTruthFeedback
+      penalize_extra_steps: false
+    dataset:
+      _target_: envs.QADatasetAdapter
+      dataset:
+        _target_: envs.RetrievalHotPotQA
+        path: ${envs.data_path}
+        split: train
+        seed: ${seed}
+  test_env:
+    _target_: envs.qa_env.QAEnv
+    max_steps: ${envs.env.max_steps}
+    max_action_length_in_memory: 256
+    action_embed_length: ${max_action_length}
+    separator: ${envs.env.separator}
+    sort_by_index: ${envs.sort_by_index}
+    positions_processor: ${envs.env.positions_processor}
+    feedback_model:
+      _target_: rl.feedback.GroundTruthFeedback
+      penalize_extra_steps: false
+    dataset:
+      _target_: envs.QADatasetAdapter
+      dataset:
+        _target_: envs.RetrievalHotPotQA
+        path: ${envs.data_path}
+        split: eval
+        seed: ${seed}
+algo:
+  model:
+    model_name: intfloat/multilingual-e5-large
+    revision: main
+    use_fast_tokenizer: true
+    predictor:
+      num_layers: 24
+      input_dim: 1024
+      hidden_dim: 512
+      output_dim: 1
+      max_seq_len: 5000
+      interpolate_factor: 1
+  action_model:
+    _target_: rl.bert_predictor.BertPredictor
+    bert:
+      _target_: transformers.AutoModel.from_pretrained
+      pretrained_model_name_or_path: ${algo.model.model_name}
+      revision: ${algo.model.revision}
+    num_hidden_layers: ${algo.model.predictor.num_layers}
+    tokenizer:
+      _target_: transformers.AutoTokenizer.from_pretrained
+      pretrained_model_name_or_path: ${algo.model.model_name}
+      revision: ${algo.model.revision}
+      use_fast: ${algo.model.use_fast_tokenizer}
+    model_dim: ${algo.model.predictor.input_dim}
+    output_size: ${algo.model.predictor.hidden_dim}
+    n_output: ${algo.model.predictor.output_dim}
+  action_embed_dict:
+    absolute:
+      _target_: rl.bert_predictor.EmbedderWithAbsoluteEncoding
+      model: ${algo.action_model}
+      max_seq_len: ${algo.model.predictor.max_seq_len}
+    random:
+      _target_: rl.bert_predictor.EmbedderWithAbsoluteEncoding
+      model: ${algo.action_model}
+      max_seq_len: ${algo.model.predictor.max_seq_len}
+      interpolate_factor: ${algo.model.predictor.interpolate_factor}
+    relative:
+      _target_: rl.bert_predictor.EmbedderWithRelativeEncoding
+      model: ${algo.action_model}
+      max_seq_len: 1000
+    none:
+      _target_: rl.bert_predictor.EmbedderNone
+      model: ${algo.action_model}
+  pqn:
+    _target_: rl.pqn.PQN
+    state_embed:
+      _target_: rl.bert_predictor.BertPredictor
+      bert:
+        _target_: transformers.AutoModel.from_pretrained
+        pretrained_model_name_or_path: ${algo.model.model_name}
+        revision: ${algo.model.revision}
+      num_hidden_layers: ${algo.model.predictor.num_layers}
+      tokenizer:
+        _target_: transformers.AutoTokenizer.from_pretrained
+        pretrained_model_name_or_path: ${algo.model.model_name}
+        revision: ${algo.model.revision}
+        use_fast: ${algo.model.use_fast_tokenizer}
+      model_dim: ${algo.model.predictor.input_dim}
+      output_size: ${algo.model.predictor.hidden_dim}
+      n_output: ${algo.model.predictor.output_dim}
+    action_embed: ${algo.action_embed_dict.${envs.positions_processor}}
+    state_embed_target:
+      _target_: rl.bert_predictor.BertPredictor
+      bert:
+        _target_: transformers.AutoModel.from_pretrained
+        pretrained_model_name_or_path: ${algo.model.model_name}
+        revision: ${algo.model.revision}
+      num_hidden_layers: ${algo.model.predictor.num_layers}
+      tokenizer:
+        _target_: transformers.AutoTokenizer.from_pretrained
+        pretrained_model_name_or_path: ${algo.model.model_name}
+        revision: ${algo.model.revision}
+        use_fast: ${algo.model.use_fast_tokenizer}
+      model_dim: ${algo.model.predictor.input_dim}
+      output_size: ${algo.model.predictor.hidden_dim}
+      n_output: ${algo.model.predictor.output_dim}
+    action_embed_target: ${algo.action_embed_dict.${envs.positions_processor}}
+    hyperparams:
+      gamma: 0.99
+      alpha: 0.05
+      Lambda: 0.5
+      tau: 0.02
+      max_grad_norm: 2.0
+      accumulate_grads: ${accumulate_grads}
+      action_embed_length: ${max_action_length}
+      max_action_length_in_memory: 256
+    optimizer:
+      _target_: torch.optim.AdamW
+      lr: 1.5e-05
+      betas:
+      - 0.9
+      - 0.98
+      eps: 1.0e-06
+      weight_decay: 0.0005
+    scheduler:
+      _target_: rl.optim.WarmupLinearScheduler
+      total: ${steps_count}
+      ratio: 0.1
+      warmup: 1000
+logger:
+  log_dir: runs/Aug24_01-52-22_PQN_hotpotqa
+  tensorboard:
+    _target_: torch.utils.tensorboard.SummaryWriter
+    comment: _PQN_${envs.task}
+    log_dir: runs/Aug24_01-52-22_PQN_hotpotqa/tb_logs/
+seed: 42
+device: cuda:0
+learning_start: 200
+steps_count: 10000
+batch_size: 12
+accumulate_grads: 8
+eval_interval: 100
+eval_episodes: 300
+envs_parallel: 1
+max_action_length: 220

qrag-ft-e5-on-hotpotqa/eval_seed42.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bed1db695951c66cac0af83507954ce822b73e64bab50ee1bf9450afff890db
+size 15601203

qrag-ft-e5-on-hotpotqa/model_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccd4068d3917bed2e44afcd7ee7a9d66c8487b44dd2206252e6d2985352f92b7
+size 20151355435