Upload folder using huggingface_hub

README.md

@@ -0,0 +1,55 @@
+---
+tags:
+- matformer
+- custom-model
+library_name: transformers
+---
+
+# Matformer Model
+
+This model was trained using the [Matformer](https://github.com/mrinaldi97/matformer) library.
+
+## Model Type
+- **Architecture**: causal
+
+## Installation
+
+First, install the required package:
+
+```bash
+pip install git+https://github.com/mrinaldi97/matformer.git
+```
+
+Or set the `MATFORMER_ROOT` environment variable:
+
+```bash
+export MATFORMER_ROOT=/path/to/matformer
+```
+
+## Usage
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# Load model
+model = AutoModelForCausalLM.from_pretrained(
+    "mrinaldi/prova001",
+    trust_remote_code=True
+)
+
+# Generate text
+prompt = "The transformer model is a"
+inputs = model.tokenizer.encode(prompt, add_bos=True, add_eos=False)
+inputs = torch.tensor([inputs], device=model.device)
+
+with torch.no_grad():
+    outputs = model.generate(inputs, max_new_tokens=50)
+    
+decoded = model.tokenizer.decode(outputs[0].tolist())
+print(decoded)
+```
+
+## Citation
+
+If you use this model, please cite the Matformer library.

config.json

@@ -0,0 +1,91 @@
+{
+  "_matformer_config_dict": {
+    "attention_type": [],
+    "bias": false,
+    "block_size_for_attention": 128,
+    "bos_token_id": 1,
+    "compile_flexattn": false,
+    "custom_layers": {},
+    "decoder": null,
+    "default_layer": {
+      "attn_impl": "flash",
+      "ffn_activation": "swiglu",
+      "hooks": {},
+      "normalization": "rmsnorm",
+      "normalization_position": "post",
+      "positional_encoding": "alibi",
+      "sliding_window_size": null
+    },
+    "encoder": null,
+    "entropy": null,
+    "eos_token_id": 2,
+    "ffn_factor": 3.0,
+    "has_entropy_model": null,
+    "has_text_autoencoder": null,
+    "hidden_size": 768,
+    "is_causal": true,
+    "mask_token_id": null,
+    "masked_substitution_rate": null,
+    "max_position_embeddings": 1024,
+    "model_class": null,
+    "name": "BabyLM",
+    "num_attention_heads": 12,
+    "num_hidden_layers": 12,
+    "pad_token_id": 0,
+    "rms_norm_eps": 1e-06,
+    "rope_theta": 10000.0,
+    "sliding_type": null,
+    "tie_word_embeddings": false,
+    "training_objective": "autoregressive",
+    "vocab_size": 32777
+  },
+  "architectures": [
+    "MatformerForCausalLM"
+  ],
+  "attention_type": [],
+  "auto_map": {
+    "AutoConfig": "modeling_matformer.MatformerConfig",
+    "AutoModel": "modeling_matformer.MatformerModel",
+    "AutoModelForCausalLM": "modeling_matformer.MatformerForCausalLM"
+  },
+  "bias": false,
+  "block_size_for_attention": 128,
+  "bos_token_id": 1,
+  "compile_flexattn": false,
+  "custom_layers": {},
+  "decoder": null,
+  "default_layer": {
+    "attn_impl": "flash",
+    "ffn_activation": "swiglu",
+    "hooks": {},
+    "normalization": "rmsnorm",
+    "normalization_position": "post",
+    "positional_encoding": "alibi",
+    "sliding_window_size": null
+  },
+  "encoder": null,
+  "entropy": null,
+  "eos_token_id": 2,
+  "ffn_factor": 3.0,
+  "has_entropy_model": null,
+  "has_text_autoencoder": null,
+  "hidden_size": 768,
+  "is_causal": true,
+  "mask_token_id": null,
+  "masked_substitution_rate": null,
+  "max_position_embeddings": 1024,
+  "model_class": null,
+  "model_type": "matformer",
+  "name": "BabyLM",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 10000.0,
+  "sliding_type": null,
+  "torch_dtype": "bfloat16",
+  "training_objective": "autoregressive",
+  "transformers_version": "4.53.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32777
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 0,
+  "transformers_version": "4.53.0.dev0"
+}

matformer_config.json

@@ -0,0 +1,65 @@
+{
+  "model_class": "Autoregressive_Model",
+  "model_config": {
+    "name": "BabyLM",
+    "hidden_size": 768,
+    "ffn_factor": 3.0,
+    "num_hidden_layers": 12,
+    "num_attention_heads": 12,
+    "vocab_size": 32777,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "pad_token_id": 0,
+    "tie_word_embeddings": false,
+    "rms_norm_eps": 1e-06,
+    "attention_type": [],
+    "max_position_embeddings": 1024,
+    "block_size_for_attention": 128,
+    "rope_theta": 10000.0,
+    "compile_flexattn": false,
+    "bias": false,
+    "training_objective": "autoregressive",
+    "is_causal": true,
+    "default_layer": {
+      "attn_impl": "flash",
+      "sliding_window_size": null,
+      "positional_encoding": "alibi",
+      "normalization": "rmsnorm",
+      "normalization_position": "post",
+      "ffn_activation": "swiglu",
+      "hooks": {}
+    },
+    "custom_layers": {}
+  },
+  "training": {
+    "optimizer": "muon",
+    "lr_scheduling": true,
+    "lr": 0.0008,
+    "final_lr": 8e-05,
+    "hold_steps": 0.013,
+    "weight_decay": 0.01,
+    "scheduler": "custom",
+    "gradient_clip_val": 1.0,
+    "warmup_steps": 0.029,
+    "max_epochs": 10,
+    "accumulate_grad_batches": 5,
+    "seed": 27,
+    "save_every_n_steps": 8000,
+    "checkpoint_name": "Baby_modello_di_prova_28L"
+  },
+  "tokenizer": {
+    "type": "huggingface",
+    "pretrained_name": "sapienzanlp/Minerva-350M-base-v1.0",
+    "varlen_strategy": "unpadding"
+  },
+  "data": {
+    "data_root": "/home/matteo/BabyLM/mdat_baby_finale/",
+    "batch_size": 36,
+    "num_workers": 1,
+    "mdat_strategy": "MinervaSpecialTokens",
+    "mdat_view": null
+  },
+  "save_dir": "./checkpoints_baby",
+  "wandb_project": "BabyLM",
+  "wandb_run_name": "Baby-Test-Model-Piccolo"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca2db7094c378d191365cd2f8ffdcdf5a745f9c15a63822f1f26cabe14b51915
+size 284831058

modeling_matformer.py

@@ -0,0 +1,53 @@
+# modeling_matformer.py
+# Auto-generated by Matformer integration for Hugging Face Hub compatibility
+import os
+import sys
+
+# Try to import matformer from environment
+matformer_root = os.getenv("MATFORMER_ROOT")
+if matformer_root:
+    matformer_root = os.path.abspath(os.path.expanduser(matformer_root))
+    if matformer_root not in sys.path:
+        sys.path.insert(0, matformer_root)
+
+try:
+    from matformer.modelling_matformer import (
+        MatformerForCausalLM,
+        MatformerForMaskedLM,
+        MatformerForSequenceClassification,
+        MatformerModel,
+        MatformerConfig,
+        register_matformer
+    )
+    register_matformer()
+except ImportError as e:
+    import subprocess
+    import tempfile
+    
+    print("Matformer not found. Attempting to install from GitHub...")
+    try:
+        with tempfile.TemporaryDirectory() as tmpdir:
+            subprocess.check_call([
+                sys.executable, "-m", "pip", "install", 
+                "git+https://github.com/mrinaldi97/matformer.git"
+            ])
+        
+        # Try importing again
+        from matformer.modelling_matformer import (
+            MatformerForCausalLM,
+            MatformerForMaskedLM,
+            MatformerForSequenceClassification,
+            MatformerModel,
+            MatformerConfig,
+            register_matformer
+        )
+        register_matformer()
+        print("Successfully installed and imported Matformer!")
+        
+    except Exception as install_error:
+        raise ImportError(
+            "Failed to install Matformer automatically. Please install manually:\n"
+            "  pip install git+https://github.com/mrinaldi97/matformer.git\n"
+            "Or set the MATFORMER_ROOT environment variable to a local clone:\n"
+            "  export MATFORMER_ROOT=/path/to/matformer"
+        ) from install_error