Update model card with TypeGround paper details

README.md

@@ -19,60 +19,45 @@ pipeline_tag: text-generation
 datasets:
   - TypeGround
   - ManyTypes4TypeScript
-widget:
-  - text: "const x = <mask>: number[];"
-    example_title: "Type Inference Example"
 ---
 
 # TypeGround_weight
 
-Fine-tuned models for TypeScript type inference, trained as part of the **TypeGround** framework. This repository contains both traditional encoder-decoder models (full fine-tune) and large language models (LoRA adapters).
+Model weights for the paper **"TypeGround: Fine-Grained Benchmarking for TypeScript Type Inference"**.
 
-## Repository Structure
+> **Abstract:** TypeScript is widely used in web applications and mini-programs, yet real-world projects still contain many variables, function parameters, and return values whose annotations are missing or inferred as coarse-grained types (e.g., `any`). Automatic type inference requires benchmarks with reliable training and evaluation signals. Existing TypeScript datasets (e.g., ManyTypes4TypeScript), however, still contain compiler-uncheckable type expressions, coarse annotations, insufficient context, and evaluation protocols that deviate from compiler semantics, which threatens the reliability of both the training labels and the evaluation signals derived from them. To mitigate the above limitation, we present TypeGround, an automatically constructed TypeScript type inference benchmark. The core idea of TypeGround is to leverage the type-checking capability of the TypeScript Compiler (TSC), together with the generative capability of large language models, to produce compiler-friendly type annotations while avoiding coarse-grained types. TypeGround covers 2,735 projects, 99,209 source files, and 1,428,997 type annotations. Experiments show that training on TypeGround improves overall EM-based MRR@5 by 23.55 percentage points on average over ManyTypes4TypeScript. The user study further confirms the reliability of TypeGround's automatic labelling process.
+>
+> **Dataset:** [github.com/fumx66/TypeGround](https://github.com/fumx66/TypeGround)
+
+## 📦 Repository Structure
 
 ```
 TypeGround_weight/
-├── CodeT5/                      # Full fine-tune
-│   ├── TypeGround/              # Trained on TypeGround dataset
-│   └── ManyTypes4TypeScrip/     # Trained on ManyTypes4TypeScript dataset
-├── CodeT5+/                     # Full fine-tune
+├── CodeT5/                      # Full fine-tune (T5ForConditionalGeneration)
+│   ├── TypeGround/
+│   └── ManyTypes4TypeScrip/
+├── CodeT5+/                     # Full fine-tune (T5ForConditionalGeneration)
 │   ├── TypeGround/
 │   └── ManyTypes4TypeScrip/
-├── UniXcoder/                   # Full fine-tune
+├── UniXcoder/                   # Full fine-tune (UniXcoder)
 │   ├── TypeGround/
 │   └── ManyTypes4TypeScrip/
-├── Llama3-8B/                   # LoRA adapters (rank=8, alpha=16)
+├── Llama3-8B/                   # LoRA adapters
 │   ├── ManyTypes4TypeScrip/lora/sft/
-│   ├── TypeGround/slice/lora/sft/
-│   └── TypeGround/snippet/lora/sft/
-├── Qwen3-14B/                   # LoRA adapters (rank=8, alpha=16)
+│   └── TypeGround/{slice,snippet}/lora/sft/
+├── Qwen3-14B/                   # LoRA adapters
 │   ├── ManyTypes4TypeScrip/lora/sft/
-│   ├── TypeGround/slice/lora/sft/
-│   └── TypeGround/snippet/lora/sft/
-├── DeepSeek-Coder-6.7B/         # LoRA adapters (rank=8, alpha=16)
+│   └── TypeGround/{slice,snippet}/lora/sft/
+├── DeepSeek-Coder-6.7B/         # LoRA adapters
 │   ├── ManyTypes4TypeScrip/lora/sft/
-│   ├── TypeGround/slice/lora/sft/
-│   └── TypeGround/snippet/lora/sft/
+│   └── TypeGround/{slice,snippet}/lora/sft/
 └── prediction.py                # Batch prediction script
 ```
 
-## Model Variants
-
-### Training Datasets
-
-| Variant | Description |
-|---|---|
-| `TypeGround` | Original TypeGround dataset |
-| `ManyTypes4TypeScript` | Extended dataset with diverse TypeScript types |
-| `slice` | Fine-tuned on sliced code contexts |
-| `snippet` | Fine-tuned on code snippets |
-
-## Usage
+## 🚀 Usage
 
 ### Traditional Models (Full Fine-tune)
 
-CodeT5, CodeT5+, and UniXcoder models can be loaded directly with Transformers:
-
 ```python
 from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
@@ -82,22 +67,16 @@ tokenizer = AutoTokenizer.from_pretrained("./CodeT5/TypeGround")
 
 ### LLMs (LoRA Adapters)
 
-LoRA adapters require loading the base model first:
-
 | Directory | Base Model |
 |---|---|
 | `Llama3-8B` | `meta-llama/Meta-Llama-3-8B-Instruct` |
 | `Qwen3-14B` | `Qwen/Qwen3-14B` |
 | `DeepSeek-Coder-6.7B` | `deepseek-ai/deepseek-coder-6.7b-instruct` |
 
-**Install vLLM:**
-
 ```bash
 pip install vllm
 ```
 
-**Serve with LoRA:**
-
 ```bash
 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
     --enable-lora \
@@ -107,22 +86,33 @@ vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
 
 ### Batch Prediction
 
-Edit the `__main__` block in `prediction.py` to point to your data and model, then run:
-
 ```bash
 python prediction.py
 ```
 
-## Citation
+## 📊 Models
+
+| Model | Architecture | Type | LoRA Config |
+|---|---|---|---|
+| CodeT5 | T5ForConditionalGeneration | Full fine-tune | — |
+| CodeT5+ | T5ForConditionalGeneration | Full fine-tune | — |
+| UniXcoder | UniXcoder | Full fine-tune | — |
+| Llama3-8B | CausalLM + LoRA | Adapter | rank=8, α=16 |
+| Qwen3-14B | CausalLM + LoRA | Adapter | rank=8, α=16 |
+| DeepSeek-Coder-6.7B | CausalLM + LoRA | Adapter | rank=8, α=16 |
+
+## 📝 Citation
 
 ```bibtex
-@dataset{typeground,
-  title={TypeGround: Fine-Grained Benchmarking for TypeScript Type Inference},
-  year={2026},
-  url={https://huggingface.co/fumx66/TypeGround_weight}
+@inproceedings{typeground,
+  title     = {TypeGround: Fine-Grained Benchmarking for TypeScript Type Inference},
+  author    = {Anonymous},
+  booktitle = {},
+  year      = {2026},
+  url       = {https://github.com/fumx66/TypeGround}
 }
 ```
 
-## License
+## 📄 License
 
 MIT License