Arctic-TILT release (#1)

- Arctic-TILT release (9ff5693aabcef672b1e3c1185aebc8050e1dd4f8)


Co-authored-by: Łukasz Borchmann <[email protected]>

README.md

@@ -0,0 +1,27 @@
+---
+license: apache-2.0
+tags:
+- text-to-text
+- document-understanding
+- document-ai
+- question-answering
+library_name: vllm
+pipeline_tag: text2text-generation
+---
+
+# Arctic-TILT
+
+**Arctic-TILT** is a state-of-the-art, sub-billion parameter model for Document Understanding. It is designed for high efficiency and long-context processing, achieving performance on par with models 1000x its size on a single 24GB GPU.
+
+* **GitHub Repository (Code & Examples):** [Snowflake-Labs/arctic-tilt](https://github.com/Snowflake-Labs/arctic-tilt)
+* **Paper:** [Arctic-TILT: Business Document Understanding at Sub-Billion Scale](https://aclanthology.org/2025.acl-industry.20/).
+
+## Model Details
+
+* **Model type:** Encoder-Decoder Transformer
+* **Parameters:** 822M
+* **Core Task:** Document AI (Question Answering, Key Information Extraction)
+
+## How to Use
+
+This model is optimized for use with `vLLM` and requires a custom scheduler and preprocessor. The `TiltPreprocessor` is available in the [GitHub repository](https://github.com/Snowflake-Labs/arctic-tilt).

config.json

@@ -0,0 +1 @@
+{"model_type": "tilt", "vocab_size": 32128, "d_model": 1024, "d_ff": 4096, "d_kv": 64, "num_attention_heads": 16, "num_hidden_layers": 24, "bias_num_buckets": 32, "bias_max_distance": 128, "bias_planar_max_distance": 100, "bias_horz_scale": 100, "bias_vert_scale": 100, "layer_norm_epsilon": 1e-06, "has_postfusion": true, "pad_token_id": 0, "eos_token_id": 1, "roi_expansion_width": 0, "unet_in_channels": 1, "unet_init_features": 32, "unet_out_features": 128, "max_seq_length": 125000, "image_limit": 250, "image_width": 768, "max_image_height": 2048, "crop_bboxes": true, "max_output_length": 128, "max_question_length": 256, "chunk_length": 1024, "prefix_separator": " : ", "answer_separator": " | "}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a94a80d2a6cb31841d95bae8ac8d5fc007fab0b1cd9dac49b9cbb43956406c3
+size 3287246704

tokenizer_config.json

@@ -0,0 +1 @@
+{"tokenizer_class": "T5TokenizerFast"}