Use custom config type to prevent Qwen3 fallback

README.md

@@ -17,7 +17,7 @@ Tiny 8M-class Qwen3-config causal LM with math-only novelty-gated GQA.
 
 | Item | Value |
 | --- | ---: |
-| Config type | `qwen3` |
+| Config type | `tinyqwen3_novelty` |
 | Parameters | 8.132M |
 | Layers | 8 |
 | Hidden size | 256 |
@@ -119,4 +119,4 @@ with torch.no_grad():
 print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
-This repo uses a self-contained remote `Qwen3Config` plus model code for a Qwen3-style dense decoder with a math-only novelty-gated attention block.
+This repo uses a self-contained remote `TinyQwen3NoveltyConfig` plus model code for a Qwen3-style dense decoder with a math-only novelty-gated attention block.

config.json

@@ -1,10 +1,10 @@
 {
-  "model_type": "qwen3",
+  "model_type": "tinyqwen3_novelty",
   "architectures": [
     "TinyQwen3NoveltyForCausalLM"
   ],
   "auto_map": {
-    "AutoConfig": "modeling_tinyqwen3_novelty.Qwen3Config",
+    "AutoConfig": "modeling_tinyqwen3_novelty.TinyQwen3NoveltyConfig",
     "AutoModelForCausalLM": "modeling_tinyqwen3_novelty.TinyQwen3NoveltyForCausalLM"
   },
   "vocab_size": 4098,

modeling_tinyqwen3_novelty.py

@@ -1,4 +1,3 @@
-
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -7,8 +6,8 @@ from transformers.generation import GenerationMixin
 from transformers.modeling_outputs import CausalLMOutput
 
 
-class Qwen3Config(PretrainedConfig):
-    model_type = "qwen3"
+class TinyQwen3NoveltyConfig(PretrainedConfig):
+    model_type = "tinyqwen3_novelty"
 
     def __init__(
         self,
@@ -174,7 +173,7 @@ class TinyQwen3NoveltyBlock(nn.Module):
 
 
 class TinyQwen3NoveltyForCausalLM(PreTrainedModel, GenerationMixin):
-    config_class = Qwen3Config
+    config_class = TinyQwen3NoveltyConfig
     base_model_prefix = ""
     _no_split_modules = ["TinyQwen3NoveltyBlock"]
     _tied_weights_keys = {}