Update model weights after training (epoch 1, loss 12.3738)

audio_decoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:69fb66efd2fcf1cca60ed861cab7e732be1f5afd1da828c8e756ecdbeaba07e4
+oid sha256:261d0afbb5e08b6b1900e3dea25eb42c412e5542bee5a4f0681a898ae9c8bcd8
 size 1458415836

audio_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12e4401d2925cec8a9b29973f98528943b9ccd70107fb8d90baacd11a897b051
+oid sha256:30c6b7e43c61588099a04e970a49bd38fd73465ef129d8e39ed9a1e8c45aeecf
 size 466150140

audio_projector.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d836099185cc6b6e01afdc72679fb120799bc0e3109a86779d04de093ee5b4f2
+oid sha256:728f8031a27e5eb3d42d6c7632872efcd0739f89bbddee2174bc9ed01776730b
 size 2099352

cross_attention.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b8d2d5d278130bf1d488e55783917dd6f62c1c5d1c30467339dba224b2ac890c
+oid sha256:27a688ad60e1a8efc783cfa66ad0ed5e9c4e74a0f5437e134ea717bdbb761eb3
 size 174191400

generator.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8e586a0ad33c4d788f5d915cda9ade7766844890c1fea0d91d514644c8b415f6
+oid sha256:3c01fd06b809de1bc14f78c1fd5f2f2cac625db3b22f0b696d532e7442aee71a
 size 629440508

llm.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:076b2a9ee201e21b04fa3113d34fb156cba0ac55c6cfbcc8c0bd79d393f00aca
+oid sha256:5f05d7774509a9338dd769956837f7b62c63ad4ab45a56fbb919230f51c876c6
 size 1506831304

modeling_xoron.py

@@ -10189,15 +10189,31 @@ class XoronModel(XoronPreTrainedModel):
                         
                         # Handle vocab size mismatch for LLM component
                         if comp_name == 'llm':
-                            # Check if embed_tokens size differs
                             embed_key = 'model.embed_tokens.weight'
+                            lm_head_key = 'lm_head.weight'
+                            
                             if embed_key in state_dict:
                                 saved_vocab_size = state_dict[embed_key].shape[0]
+                                hidden_size = state_dict[embed_key].shape[1]
                                 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)
+                                    # Manually resize embed_tokens
+                                    new_embed = nn.Embedding(saved_vocab_size, hidden_size)
+                                    new_embed.weight.data = state_dict[embed_key]
+                                    component.model.embed_tokens = new_embed
+                                    
+                                    # Manually resize lm_head if present
+                                    if lm_head_key in state_dict:
+                                        new_lm_head = nn.Linear(hidden_size, saved_vocab_size, bias=False)
+                                        new_lm_head.weight.data = state_dict[lm_head_key]
+                                        component.lm_head = new_lm_head
+                                    
+                                    # Remove these keys from state_dict since we already loaded them
+                                    del state_dict[embed_key]
+                                    if lm_head_key in state_dict:
+                                        del state_dict[lm_head_key]
                         
                         component.load_state_dict(state_dict, strict=False)
                         print(f"   ✅ Loaded {comp_name}")
@@ -11105,15 +11121,31 @@ class XoronModel(XoronPreTrainedModel):
                         
                         # Handle vocab size mismatch for LLM component
                         if comp_name == 'llm':
-                            # Check if embed_tokens size differs
                             embed_key = 'model.embed_tokens.weight'
+                            lm_head_key = 'lm_head.weight'
+                            
                             if embed_key in state_dict:
                                 saved_vocab_size = state_dict[embed_key].shape[0]
+                                hidden_size = state_dict[embed_key].shape[1]
                                 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)
+                                    # Manually resize embed_tokens
+                                    new_embed = nn.Embedding(saved_vocab_size, hidden_size)
+                                    new_embed.weight.data = state_dict[embed_key]
+                                    component.model.embed_tokens = new_embed
+                                    
+                                    # Manually resize lm_head if present
+                                    if lm_head_key in state_dict:
+                                        new_lm_head = nn.Linear(hidden_size, saved_vocab_size, bias=False)
+                                        new_lm_head.weight.data = state_dict[lm_head_key]
+                                        component.lm_head = new_lm_head
+                                    
+                                    # Remove these keys from state_dict since we already loaded them
+                                    del state_dict[embed_key]
+                                    if lm_head_key in state_dict:
+                                        del state_dict[lm_head_key]
                         
                         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:fa7c80f806f46f1616d3acfe6048b56e597e36027b7700eefd9ca13e5b868da9
+oid sha256:899fc13cba54ecbf1eef6401108393ed0e9b8e4584e8d180947f87d4807c39dd
 size 52880664

trainer_state.json

@@ -1,6 +1,6 @@
 {
   "best_model_checkpoint": "/kaggle/working/xoron-final",
-  "best_metric": 12.640625,
+  "best_metric": 12.373827934265137,
   "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:b9e348eb62b15d8fe7418f5c527b62ddecd4e918c36b6ebce28a035479b9432e
+oid sha256:09c47de74291bae883f60d4986fe8dbc38a2c68de7574fdabc66ce46222ed711
 size 5143

video_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f72e26338eb6855e72508d1a57a3d038d63c84ab92a9448307ad5b3430393bca
+oid sha256:f70226e533706675adf13f72c46122854021d13fe388445bc4d6b7495fa64e3a
 size 1923089112

video_generator.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1aa2bef313e63c8c8dbead2f22c6e044abd56883629f7504e54953e37e62f00b
+oid sha256:c68805a467c37a4b172786a99fb83589c8a4e40b7b7a40886176cf1fd2188dc5
 size 61574134

vision_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bc7028344414ab330bb1b2d321f2dd2e828f77a13da8b68c038a4790e170df2b
+oid sha256:4ba3ed68a4082c90661c88d125f4dd6b40717652173dbdc31df0c8cc5fa6260a
 size 1000535480

waveform_decoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9bff6c300b238aa47a1673e2ffd1f65c029916e9271cd0952b3d6e873b11837a
+oid sha256:7416892f4786903c832800ef5baac11cf3a787979cf223afa39a00434b5c639f
 size 34681076