Upload model

config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "embedding-model-mistral-64m-contrastive/checkpoint-7750",
+  "architectures": [
+    "MistralModelEmbedding"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoModel": "mistral_contrastive.MistralModelEmbedding"
+  },
+  "bos_token_id": 1,
+  "embedding_size": 768,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 512,
+  "initializer_range": 0.02,
+  "intermediate_size": 2048,
+  "max_position_embeddings": 32768,
+  "model_type": "mistral",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 8,
+  "num_key_value_heads": 8,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-05,
+  "rope_theta": 10000.0,
+  "sliding_window": 4096,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.38.1",
+  "use_cache": true,
+  "vocab_size": 32000
+}

mistral_contrastive.py

@@ -0,0 +1,60 @@
+
+from transformers import MistralPreTrainedModel, MistralModel, MistralConfig
+from typing import Dict
+from transformers.file_utils import ModelOutput
+from typing import List, Optional, Tuple, Union
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from torch import nn, Tensor
+from dataclasses import dataclass
+from torch import nn
+import torch
+from transformers.file_utils import ModelOutput
+import torch.nn.functional as F
+
+COSINE_DISTANCE = lambda x, y: 1-F.cosine_similarity(x, y)
+
+@dataclass
+class EncoderOutput(ModelOutput):
+    loss: Optional[Tensor] = None
+
+class MistralModelEmbedding(MistralPreTrainedModel):
+    def __init__(self, config, **kwargs):
+        super().__init__(config, **kwargs)
+
+        self.model = MistralModel(config)
+        self.dense_layer = nn.Linear(
+            self.config.hidden_size,
+            self.config.embedding_size,
+            bias=False
+        )
+        self.post_init()
+    
+
+    def encode(self, features):
+        if features is None:
+            return None
+        psg_out = self.model.forward(**features,return_dict=True)
+        logits = self.dense_layer(psg_out.last_hidden_state)
+        input_ids = features['input_ids']
+        batch_size = input_ids.shape[0]
+        sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+        sequence_lengths = sequence_lengths % input_ids.shape[-1]
+        sequence_lengths = sequence_lengths.to(logits.device)
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+        return pooled_logits
+
+
+    def forward(self, query: Dict[str, Tensor] = None,
+                passage: Dict[str, Tensor] = None, labels = None, margin = 1.0):
+        q_reps = self.encode(query)
+        p_reps = self.encode(passage)
+
+        loss = None
+        if labels is not None:
+            distances = COSINE_DISTANCE(q_reps, p_reps)
+            losses = 0.5 * (labels.float() * distances.pow(2) + (1 - labels).float() * F.relu(margin - distances).pow(2))
+            loss = losses.mean()
+
+        return EncoderOutput(
+            loss=loss,
+        )

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8779c25f343a3e09f5602e3e32991a6e7db339d547adc5b4091113068d6052f6
+size 96494632