Update model weights after training (epoch 1, loss 12.7932)

audio_decoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:33cd64538ab48c690ed0afbc044a236d26f6b7a731b56366e6cc3da4d2b1bb1a
+oid sha256:fdd54c3025809382728dce3459184584a22eddee107472efc179a52b01e60409
 size 1458415836

audio_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cd06846915bf81f9619f0f235bc87432af972f108fabd5641c6fedf4bc211d5f
+oid sha256:f38f6922edf7291ac97e6bd509830bd202978d1dac41f57e11f891753367de4d
 size 466150140

audio_projector.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0f14ee94b2dbc12d2bcb7122f1c007602419a4ef10b92b724ede52fdd2e488a0
+oid sha256:27d2510596f713b85ddc3d6e172277b34c6ef824a5632058496d67093506863d
 size 2099352

cross_attention.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d81f0f0b7fae83ed9cc582826729c09c340665f8e5bfa11fa2a3a1b34fcc3b81
+oid sha256:78710f2a29bf0a2d75cffef87817d3fe899330004d44861541e5d3b0381a3ed2
 size 174191400

generator.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6abe298e9826a4652f91afa17a3393657eb6267d866bd8bde437cdf44186b9d
+oid sha256:a194896fc7c1c35d6013adf3dc3bb1e18b01dfad74b1cedb2fe13670b65ccedd
 size 629440508

llm.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:64f0666b62452a957a9caa7b14ae17194cc7fe73f03ae7131b2cd1ba0448cc9f
+oid sha256:5f099a9a521f89605c1c237f9ce03aa9d9b5a204cac5cffee72a39adbfe13fbd
 size 1506831304

modeling_xoron.py

@@ -475,9 +475,8 @@ class XoronModel(XoronPreTrainedModel):
         """
         Save model to directory in HuggingFace format.
         
-        This saves both the model weights and dynamically builds a self-contained
-        modeling_xoron.py that can be loaded with trust_remote_code=True without
-        requiring the full Xoron-Dev package.
+        This saves both the model weights and the custom code files
+        needed for trust_remote_code loading.
         """
         os.makedirs(save_directory, exist_ok=True)
         
@@ -488,18 +487,21 @@ class XoronModel(XoronPreTrainedModel):
         if self._internal_model is not None:
             self._internal_model.save_pretrained(save_directory)
         
-        # Copy configuration file
+        # Copy custom code files for trust_remote_code
         import shutil
         current_dir = os.path.dirname(os.path.abspath(__file__))
         
-        config_src = os.path.join(current_dir, 'configuration_xoron.py')
-        config_dst = os.path.join(save_directory, 'configuration_xoron.py')
-        if os.path.exists(config_src):
-            shutil.copy2(config_src, config_dst)
+        # Files to copy
+        files_to_copy = [
+            'configuration_xoron.py',
+            'modeling_xoron.py',
+        ]
         
-        # Build self-contained modeling_xoron.py dynamically
-        modeling_dst = os.path.join(save_directory, 'modeling_xoron.py')
-        self._build_self_contained_modeling_file(current_dir, modeling_dst)
+        for filename in files_to_copy:
+            src = os.path.join(current_dir, filename)
+            dst = os.path.join(save_directory, filename)
+            if os.path.exists(src):
+                shutil.copy2(src, dst)
         
         # Update config.json with auto_map for trust_remote_code
         config_path = os.path.join(save_directory, 'config.json')
@@ -518,284 +520,6 @@ class XoronModel(XoronPreTrainedModel):
         
         if push_to_hub:
             self.push_to_hub(save_directory, **kwargs)
-    
-    def _build_self_contained_modeling_file(self, repo_root: str, output_path: str):
-        """
-        Build a self-contained modeling_xoron.py by combining all model components.
-        
-        This allows HuggingFace users to load the model with trust_remote_code=True
-        without needing to install the full Xoron-Dev package.
-        """
-        import re
-        
-        # Component files to combine (in dependency order)
-        component_files = [
-            "models/components/lora.py",
-            "models/components/attention.py",
-            "models/components/projectors.py",
-            "models/components/moe.py",
-            "models/encoders/vision.py",
-            "models/encoders/video.py",
-            "models/encoders/audio.py",
-            "models/generators/image.py",
-            "models/generators/video.py",
-            "models/llm/moe_llama.py",
-            "models/xoron.py",
-        ]
-        
-        # Internal imports to remove
-        internal_import_patterns = [
-            r"^from config import.*$",
-            r"^from config\..*import.*$",
-            r"^from models\..*import.*$",
-            r"^from models import.*$",
-        ]
-        
-        def is_internal_import(line):
-            line = line.strip()
-            for pattern in internal_import_patterns:
-                if re.match(pattern, line):
-                    return True
-            return False
-        
-        def is_external_import(line):
-            line = line.strip()
-            return (line.startswith("import ") or 
-                    (line.startswith("from ") and not is_internal_import(line)))
-        
-        def extract_code_body(content):
-            """Extract code body, removing module docstring and imports."""
-            lines = content.split('\n')
-            code_lines = []
-            i = 0
-            
-            # Skip leading whitespace
-            while i < len(lines) and not lines[i].strip():
-                i += 1
-            
-            # Skip module docstring if present
-            if i < len(lines):
-                stripped = lines[i].strip()
-                if stripped.startswith('"""') or stripped.startswith("'''"):
-                    docstring_char = stripped[:3]
-                    if stripped.count(docstring_char) >= 2:
-                        i += 1
-                    else:
-                        i += 1
-                        while i < len(lines):
-                            if docstring_char in lines[i]:
-                                i += 1
-                                break
-                            i += 1
-            
-            # Process remaining lines
-            for line in lines[i:]:
-                stripped = line.strip()
-                
-                # Skip empty lines at start
-                if not code_lines and not stripped:
-                    continue
-                
-                # Skip internal imports
-                if is_internal_import(line):
-                    continue
-                
-                # Skip external imports (we'll add our own)
-                if is_external_import(line):
-                    continue
-                
-                # Skip logger setup lines
-                if stripped.startswith("logger = logging.getLogger"):
-                    continue
-                
-                code_lines.append(line)
-            
-            # Remove trailing empty lines
-            while code_lines and not code_lines[-1].strip():
-                code_lines.pop()
-            
-            return '\n'.join(code_lines)
-        
-        # Build the file
-        header = '''"""
-Xoron Model for HuggingFace Transformers - Self-Contained Implementation.
-
-AUTO-GENERATED FILE - Do not edit directly!
-
-This module provides a complete, self-contained HuggingFace-compatible model class
-for the Xoron multimodal model. All components are embedded directly in this file
-to enable loading via AutoModel with trust_remote_code=True WITHOUT requiring
-the full Xoron-Dev package to be installed.
-
-Usage:
-    from transformers import AutoModel, AutoConfig
-    config = AutoConfig.from_pretrained("your-repo/xoron-model", trust_remote_code=True)
-    model = AutoModel.from_pretrained("your-repo/xoron-model", trust_remote_code=True)
-"""
-
-import os
-import math
-import json
-import logging
-from dataclasses import dataclass, field
-from typing import Optional, Dict, List, Union, Tuple, Any
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-try:
-    from safetensors.torch import save_model, load_model
-except ImportError:
-    save_model, load_model = None, None
-
-from transformers import PreTrainedModel, LlamaConfig, LlamaModel, LlamaForCausalLM
-from transformers.modeling_outputs import CausalLMOutputWithPast, BaseModelOutputWithPast
-
-try:
-    from transformers.models.llama.modeling_llama import (
-        LlamaAttention, LlamaDecoderLayer, LlamaRMSNorm, LlamaMLP,
-        LlamaRotaryEmbedding, apply_rotary_pos_emb, repeat_kv
-    )
-except ImportError:
-    LlamaAttention = LlamaDecoderLayer = LlamaRMSNorm = LlamaMLP = None
-    LlamaRotaryEmbedding = apply_rotary_pos_emb = repeat_kv = None
-
-# Import configuration
-try:
-    from .configuration_xoron import XoronConfig
-except ImportError:
-    from configuration_xoron import XoronConfig
-
-logger = logging.getLogger(__name__)
-
-'''
-        
-        all_code = [header]
-        
-        # Process each component file
-        for filepath in component_files:
-            full_path = os.path.join(repo_root, filepath)
-            if not os.path.exists(full_path):
-                continue
-            
-            with open(full_path, 'r', encoding='utf-8') as f:
-                content = f.read()
-            
-            code = extract_code_body(content)
-            if code.strip():
-                section_name = filepath.replace('/', '.').replace('.py', '').upper()
-                section_header = f"\n\n# {'='*78}\n# {section_name}\n# {'='*78}\n\n"
-                all_code.append(section_header + code)
-        
-        # Add HuggingFace wrapper classes
-        hf_wrapper = '''
-
-# ==============================================================================
-# HUGGINGFACE WRAPPER CLASSES
-# ==============================================================================
-
-class XoronPreTrainedModel(PreTrainedModel):
-    """Base class for Xoron models providing HuggingFace integration."""
-    
-    config_class = XoronConfig
-    base_model_prefix = "xoron"
-    supports_gradient_checkpointing = True
-    _no_split_modules = ["XoronMultimodalModel"]
-    _skip_keys_device_placement = "past_key_values"
-    
-    def _init_weights(self, module):
-        std = 0.02
-        if isinstance(module, nn.Linear):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.bias is not None:
-                module.bias.data.zero_()
-        elif isinstance(module, nn.Embedding):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.padding_idx is not None:
-                module.weight.data[module.padding_idx].zero_()
-        elif isinstance(module, nn.LayerNorm):
-            module.bias.data.zero_()
-            module.weight.data.fill_(1.0)
-
-
-class XoronModel(XoronPreTrainedModel):
-    """Xoron Multimodal Model for HuggingFace."""
-    
-    def __init__(self, config: XoronConfig):
-        super().__init__(config)
-        self.config = config
-        self._internal_model = XoronMultimodalModel(config)
-        self.post_init()
-    
-    @property
-    def internal_model(self):
-        return self._internal_model
-    
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        pixel_values: Optional[torch.Tensor] = None,
-        video_frames: Optional[torch.Tensor] = None,
-        audio_features: Optional[torch.Tensor] = None,
-        labels: Optional[torch.Tensor] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = None,
-        **kwargs,
-    ) -> Union[Tuple, CausalLMOutputWithPast]:
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-        
-        outputs = self._internal_model(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            images=pixel_values,
-            video=video_frames,
-            audio=audio_features,
-            labels=labels,
-        )
-        
-        if return_dict:
-            return CausalLMOutputWithPast(
-                loss=outputs.get('loss'),
-                logits=outputs.get('logits'),
-                past_key_values=outputs.get('past_key_values'),
-                hidden_states=outputs.get('hidden_states'),
-                attentions=outputs.get('attentions'),
-            )
-        
-        return (outputs.get('loss'), outputs.get('logits'))
-    
-    def generate_image(self, prompt_embeds: torch.Tensor, **kwargs):
-        return self._internal_model.generate_image(prompt_embeds, **kwargs)
-    
-    def generate_video(self, prompt_embeds: torch.Tensor, **kwargs):
-        return self._internal_model.generate_video(prompt_embeds, **kwargs)
-    
-    def generate_speech(self, text_embeds: torch.Tensor, **kwargs):
-        return self._internal_model.generate_speech(text_embeds, **kwargs)
-
-
-class XoronForCausalLM(XoronModel):
-    """Alias for XoronModel for compatibility."""
-    pass
-
-
-# Register for AutoClass
-XoronConfig.register_for_auto_class()
-XoronModel.register_for_auto_class("AutoModel")
-XoronForCausalLM.register_for_auto_class("AutoModelForCausalLM")
-'''
-        all_code.append(hf_wrapper)
-        
-        # Write output file
-        final_content = '\n'.join(all_code)
-        with open(output_path, 'w', encoding='utf-8') as f:
-            f.write(final_content)
-        
-        line_count = final_content.count('\n')
-        print(f"   📦 Built self-contained modeling_xoron.py ({line_count:,} lines)")
 
 
 class XoronForCausalLM(XoronModel):

projector.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:29201dd7bea97c27d99d71cb06a71f1c8122ff185b7833957b49627709c1a2c8
+oid sha256:7f70e42cfac214c75c189d74f0721c19a0db5d0da465ec8e79dc90a1a0be94bb
 size 52880664

streaming_state.json

@@ -3,20 +3,20 @@
   "unique_samples": 1,
   "total_yields": 2,
   "dataset_positions": {
-    "WebSight": 1
+    "Sora-Style-Likert": 1
   },
   "modality_positions": {
     "text": {},
-    "image": {
-      "WebSight": 1
+    "image": {},
+    "video": {
+      "Sora-Style-Likert": 1
     },
-    "video": {},
     "audio": {}
   },
   "modality_counts": {
     "text": 0,
-    "image": 1,
-    "video": 0,
+    "image": 0,
+    "video": 1,
     "audio": 0
   },
   "last_modality": null

trainer_state.json

@@ -1,6 +1,6 @@
 {
   "best_model_checkpoint": "/kaggle/working/xoron-final",
-  "best_metric": 12.430816173553467,
+  "best_metric": 12.793164253234863,
   "epoch": 1,
   "epochs_completed": 1,
   "global_step": 0,
@@ -17,16 +17,16 @@
   "max_grad_norm": 1.0,
   "trainable_components": [
     "vision",
+    "video",
     "llm",
     "cross_attention",
-    "image_generation",
+    "video_generation",
     "modality_markers"
   ],
   "frozen_components": [
-    "video",
     "audio",
     "speech",
-    "video_generation"
+    "image_generation"
   ],
   "trial_name": null,
   "trial_params": null

training_state.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cabf5d817b134ab53a234f26371f536f37e818b23811dcd0c6142c335f71166f
-size 5143
+oid sha256:8ed0a7799b91727f0ad203a1b686176319c5556cb441238bde744d765824dc67
+size 6167

video_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91a62a45d920fd3989d728c1cbb851cea12c081330aaaefb58a6d78d39aa4c63
+oid sha256:4ca3d1114993b0267e8f769779c8774b22a3f0bba781436311c292c745d7a4a6
 size 1923089112

video_generator.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:775996f8c250140d886f7c03673370969f829088c7796143d291796fdc2b16ac
+oid sha256:0138141f5e37860df2a57152882d24962e042f0781249edd16b3b2934e62c178
 size 61574134

vision_encoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:49c149f367899d37da59699a65c41fd0a8c8aafd757c09491b565f751279318c
+oid sha256:db8657fa0918d538c717c153d0104d49d5cc7417d34c1fd36cf8d278e653f9df
 size 1000535480

waveform_decoder.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ec52685af5d97e3adf34b326c8ec2b821ee9b4ed084249953afae524f056019a
+oid sha256:9778a8e34136fe9dd2603100b3a20da0b80423573f89068bac005e9eef239fcb
 size 34681076