Update model weights after training (epoch 1, loss 12.6406)

audio_decoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:345c19f21b1d9f02ad3054d3cc7e9e48ea5e82c276334869646391d4b0596e79
+oid sha256:69fb66efd2fcf1cca60ed861cab7e732be1f5afd1da828c8e756ecdbeaba07e4
 size 1458415836

audio_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:258d79b8e1b755907bfd5083947451386aa41f8370b4f56c883ba9e9f54f46ba
+oid sha256:12e4401d2925cec8a9b29973f98528943b9ccd70107fb8d90baacd11a897b051
 size 466150140

audio_projector.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c48261a28a843318cfc7afa7e5ff9e70a997f465b78ea8be64509c42861c56eb
+oid sha256:d836099185cc6b6e01afdc72679fb120799bc0e3109a86779d04de093ee5b4f2
 size 2099352

cross_attention.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d183ddac32911ec28151d1569c00bd5804910fc622fd35853ed661b7ffa4f5a3
+oid sha256:b8d2d5d278130bf1d488e55783917dd6f62c1c5d1c30467339dba224b2ac890c
 size 174191400

generator.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:83d86a5edad6600d32976ad9dbce1b2202d40f42429d1bcd44484f00f6caf6b3
+oid sha256:8e586a0ad33c4d788f5d915cda9ade7766844890c1fea0d91d514644c8b415f6
 size 629440508

llm.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a1a06cd599c200f1965030fd038048e2437decebe316d5ad0167d2908b217128
+oid sha256:076b2a9ee201e21b04fa3113d34fb156cba0ac55c6cfbcc8c0bd79d393f00aca
 size 1506831304

modeling_xoron.py

@@ -10186,6 +10186,19 @@ class XoronModel(XoronPreTrainedModel):
                     if component is not None:
                         with safe_open(comp_path, framework="pt") as f:
                             state_dict = {k: f.get_tensor(k) for k in f.keys()}
+                        
+                        # Handle vocab size mismatch for LLM component
+                        if comp_name == 'llm':
+                            # Check if embed_tokens size differs
+                            embed_key = 'model.embed_tokens.weight'
+                            if embed_key in state_dict:
+                                saved_vocab_size = state_dict[embed_key].shape[0]
+                                current_vocab_size = component.model.embed_tokens.weight.shape[0]
+                                
+                                if saved_vocab_size != current_vocab_size:
+                                    print(f"   📐 Resizing embeddings: {current_vocab_size} -> {saved_vocab_size}")
+                                    component.resize_token_embeddings(saved_vocab_size)
+                        
                         component.load_state_dict(state_dict, strict=False)
                         print(f"   ✅ Loaded {comp_name}")
             
@@ -11089,6 +11102,19 @@ class XoronModel(XoronPreTrainedModel):
                     if component is not None:
                         with safe_open(comp_path, framework="pt") as f:
                             state_dict = {k: f.get_tensor(k) for k in f.keys()}
+                        
+                        # Handle vocab size mismatch for LLM component
+                        if comp_name == 'llm':
+                            # Check if embed_tokens size differs
+                            embed_key = 'model.embed_tokens.weight'
+                            if embed_key in state_dict:
+                                saved_vocab_size = state_dict[embed_key].shape[0]
+                                current_vocab_size = component.model.embed_tokens.weight.shape[0]
+                                
+                                if saved_vocab_size != current_vocab_size:
+                                    print(f"   📐 Resizing embeddings: {current_vocab_size} -> {saved_vocab_size}")
+                                    component.resize_token_embeddings(saved_vocab_size)
+                        
                         component.load_state_dict(state_dict, strict=False)
                         print(f"   ✅ Loaded {comp_name}")
             

projector.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:42848d798e1fd88408f15cb242e3e117a6dc1166a68d85bd2b60d556902b3567
+oid sha256:fa7c80f806f46f1616d3acfe6048b56e597e36027b7700eefd9ca13e5b868da9
 size 52880664

streaming_state.json

@@ -3,12 +3,12 @@
   "unique_samples": 1,
   "total_yields": 2,
   "dataset_positions": {
-    "Football": 1
+    "InstructPix2Pix": 1
   },
   "modality_positions": {
     "text": {},
     "image": {
-      "Football": 1
+      "InstructPix2Pix": 1
     },
     "video": {},
     "audio": {}

trainer_state.json

@@ -1,6 +1,6 @@
 {
   "best_model_checkpoint": "/kaggle/working/xoron-final",
-  "best_metric": 12.606770992279053,
+  "best_metric": 12.640625,
   "epoch": 1,
   "epochs_completed": 1,
   "global_step": 0,

training_state.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:486d68f890c9804dd2c27b25f84601ad3fd4dab4089d237c2e7ffb37883a9a7b
+oid sha256:b9e348eb62b15d8fe7418f5c527b62ddecd4e918c36b6ebce28a035479b9432e
 size 5143

video_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9c65e415613ad0a88b9d0a4f7b393d21f138a6f5be3aeb89bc26c010a028a447
+oid sha256:f72e26338eb6855e72508d1a57a3d038d63c84ab92a9448307ad5b3430393bca
 size 1923089112

video_generator.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f86358c587f5dd7e17c5cd17898de9ffe1104fba7b7cba7a7ae2e4a54e3a9a14
+oid sha256:1aa2bef313e63c8c8dbead2f22c6e044abd56883629f7504e54953e37e62f00b
 size 61574134

vision_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a467be0cf69c30b067287b20aa89f9031a50dbe4106938a64aa87e922ad2d096
+oid sha256:bc7028344414ab330bb1b2d321f2dd2e828f77a13da8b68c038a4790e170df2b
 size 1000535480

waveform_decoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7408a6ce0c03a47f12b4f4f3a2222f04cd7b3651dfcc1335aaa8e2648fddb029
+oid sha256:9bff6c300b238aa47a1673e2ffd1f65c029916e9271cd0952b3d6e873b11837a
 size 34681076