Training in progress, step 50

README.md

@@ -1,17 +1,17 @@
 ---
-base_model: LiquidAI/LFM2-VL-1.6B
+base_model: LiquidAI/LFM2-VL-450M
 library_name: transformers
 model_name: lfm2-vl-iconclass
 tags:
 - generated_from_trainer
-- trl
 - sft
+- trl
 licence: license
 ---
 
 # Model Card for lfm2-vl-iconclass
 
-This model is a fine-tuned version of [LiquidAI/LFM2-VL-1.6B](https://huggingface.co/LiquidAI/LFM2-VL-1.6B).
+This model is a fine-tuned version of [LiquidAI/LFM2-VL-450M](https://huggingface.co/LiquidAI/LFM2-VL-450M).
 It has been trained using [TRL](https://github.com/huggingface/trl).
 
 ## Quick start
@@ -27,7 +27,7 @@ print(output["generated_text"])
 
 ## Training procedure
 
-[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/davanstrien/huggingface/runs/9vfvugcn) 
+[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/davanstrien/huggingface/runs/0da6dcsh) 
 
 
 This model was trained with SFT.

config.json

@@ -12,7 +12,7 @@
   "image_token_index": 396,
   "max_image_tokens": 256,
   "max_num_patches": 1024,
-  "max_pixels_tolerance": 1.5,
+  "max_pixels_tolerance": 2.0,
   "max_tiles": 10,
   "min_image_tokens": 64,
   "min_tiles": 2,
@@ -21,13 +21,13 @@
   "projector_hidden_act": "gelu",
   "projector_hidden_size": 2560,
   "text_config": {
-    "_name_or_path": "LiquidAI/LFM2-1.2B",
+    "_name_or_path": "LiquidAI/LFM2-350M",
     "architectures": [
       "Lfm2ForCausalLM"
     ],
     "block_auto_adjust_ff_dim": true,
-    "block_dim": 2048,
-    "block_ff_dim": 12288,
+    "block_dim": 1024,
+    "block_ff_dim": 6656,
     "block_ffn_dim_multiplier": 1.0,
     "block_mlp_init_scale": 1.0,
     "block_multiple_of": 256,
@@ -37,13 +37,13 @@
     "block_use_xavier_init": true,
     "conv_L_cache": 3,
     "conv_bias": false,
-    "conv_dim": 2048,
-    "conv_dim_out": 2048,
+    "conv_dim": 1024,
+    "conv_dim_out": 1024,
     "conv_use_xavier_init": true,
     "eos_token_id": 7,
-    "hidden_size": 2048,
+    "hidden_size": 1024,
     "initializer_range": 0.02,
-    "intermediate_size": 12288,
+    "intermediate_size": 6656,
     "layer_types": [
       "conv",
       "conv",
@@ -65,8 +65,8 @@
     "max_position_embeddings": 128000,
     "model_type": "lfm2",
     "norm_eps": 1e-05,
-    "num_attention_heads": 32,
-    "num_heads": 32,
+    "num_attention_heads": 16,
+    "num_heads": 16,
     "num_hidden_layers": 16,
     "num_key_value_heads": 8,
     "rope_theta": 1000000.0,
@@ -79,21 +79,21 @@
   "torch_dtype": "bfloat16",
   "transformers_version": "4.55.0",
   "use_image_special_tokens": true,
-  "use_thumbnail": true,
+  "use_thumbnail": false,
   "vision_config": {
     "attention_dropout": 0.0,
     "hidden_act": "gelu_pytorch_tanh",
-    "hidden_size": 1152,
-    "intermediate_size": 4304,
+    "hidden_size": 768,
+    "intermediate_size": 3072,
     "layer_norm_eps": 1e-06,
     "model_type": "siglip2_vision_model",
-    "num_attention_heads": 16,
+    "num_attention_heads": 12,
     "num_channels": 3,
-    "num_hidden_layers": 27,
+    "num_hidden_layers": 12,
     "num_patches": 256,
     "patch_size": 16,
     "torch_dtype": "bfloat16",
     "vision_use_head": false
   },
-  "vision_feature_layer": -2
+  "vision_feature_layer": -1
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ae00d8b556f6fac6ab9f616bd576008cc2d349d5fa02e497ff899a0b2b4d4351
-size 3169688328
+oid sha256:29b5cf0aa9e685c13492826cbb781caef44eeab998657271d12e2ea33297f169
+size 901692416

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:63bf6315f8bebb183d9c173d0795709f60f13031db9034ecf9c1a14bb3c849d4
+oid sha256:c049c52fd40bedd7b5fabd7c160997c99db0c09fef4e049b41c801ff07e6dd0d
 size 6161