Upload model

config.json

@@ -0,0 +1,40 @@
+{
+  "_name_or_path": "embedding-model-llama-2b-contrastive/checkpoint-8100",
+  "architectures": [
+    "LlamaModelEmbedding"
+  ],
+  "attention_bias": false,
+  "auto_map": {
+    "AutoModel": "modeling_contrastive.LlamaModelEmbedding"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "id2label": {
+    "0": "LABEL_0"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "label2id": {
+    "LABEL_0": 0
+  },
+  "max_position_embeddings": 32768,
+  "model_type": "llama",
+  "normalized": true,
+  "num_attention_heads": 40,
+  "num_hidden_layers": 5,
+  "num_key_value_heads": 40,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "sentence_pooling_method": "mean",
+  "temperature": 0.02,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.35.2",
+  "use_cache": true,
+  "vocab_size": 32000
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa5fda645b43bf269f5d548ccfe9dddbbc17e98c06901f1d3366ef782ecae451
+size 3515472056

modeling_contrastive.py

@@ -0,0 +1,60 @@
+
+from transformers import LlamaModel, LlamaConfig, LlamaTokenizer
+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
+from typing import Dict
+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 LlamaModelEmbedding(LlamaModel):
+    def __init__(self, config: LlamaConfig, **kwargs):
+        super().__init__(config, **kwargs)
+
+        self.dense_layer = nn.Linear(self.config.hidden_size,1536)
+        
+    def sentence_embedding(self, hidden_state, mask):
+        if self.config.sentence_pooling_method == 'mean':
+            s = torch.sum(hidden_state * mask.unsqueeze(-1).float(), dim=1)
+            d = mask.sum(axis=1, keepdim=True).float()
+            return s / d
+        elif self.config.sentence_pooling_method == 'cls':
+            return hidden_state[:,0]
+
+    def encode(self, features):
+        if features is None:
+            return None
+        psg_out = super().forward(**features,return_dict=True)
+        output = self.dense_layer(psg_out.last_hidden_state)
+        p_reps = self.sentence_embedding(output, features['attention_mask'])
+        if self.config.normalized:
+            p_reps = torch.nn.functional.normalize(p_reps, dim=-1)
+        return p_reps.contiguous()
+
+
+    def forward(self, query: Dict[str, Tensor] = None,
+                passage: Dict[str, Tensor] = None, labels = None, margin = 0.5):
+        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,
+        )
+