docs: 添加英文版模型卡 README_en.md

- 保留 Hugging Face YAML 元数据
- 添加完整的英文模型说明
- 包含使用方法、评测结果、数据集引用
- 与中文版 README.md 内容对齐

README_en.md

@@ -4,6 +4,7 @@ datasets:
 - pjura/mahjong_board_states
 language:
 - zh
+- en
 base_model:
 - unsloth/Qwen3-4B-Instruct-2507
 tags:
@@ -19,6 +20,8 @@ pipeline_tag: text-generation
 
 # Qwen3-4B-Instruct-2507-mahjong-alpha
 
+[中文](./README.md)
+
 `Qwen3-4B-Instruct-2507-mahjong-alpha` is a Riichi Mahjong domain model fine-tuned from `unsloth/Qwen3-4B-Instruct-2507` with QLoRA.
 
 It is designed for 4-player Riichi Mahjong discard recommendation: given round information, hand tiles, calls, visible tiles, tile-efficiency, and defense signals, the model outputs the single best discard tile for the current state.
@@ -28,10 +31,10 @@ The current version is mainly intended for tool integration. The output is a sin
 ## Model Features
 
 - **Task**: 4-player Riichi Mahjong discard recommendation
-- **Base Model**: `unsloth/Qwen3-4B-Instruct-2507`
+- **Base model**: `unsloth/Qwen3-4B-Instruct-2507`
 - **Fine-tuning**: `QLoRA`
-- **Training Framework**: `Unsloth`
-- **Release Format**: `GGUF (F16)`
+- **Training framework**: `Unsloth`
+- **Release format**: `GGUF (F16)`
 - **Inference**: `llama.cpp`
 - **Maintainer**: `TTDXQ`
 
@@ -90,9 +93,9 @@ The output is strictly a single tile text without any prefix like "discard" and
 白
 ```
 
-## How to Use
+## Usage
 
-### Inference with llama.cpp
+### llama.cpp Inference
 
 ```bash
 llama-server -m Qwen3-4B-Instruct-2507-mahjong-alpha.gguf -c 2048
@@ -111,7 +114,10 @@ tokenizer = AutoTokenizer.from_pretrained(
 )
 
 # Prepare input
-input_text = "[情景分析]\n- 牌局: 东一局，你是庄家 (第1巡，牌墙余69张)。\n..."
+input_text = """[情景分析]
+- 牌局: 东一局，你是庄家 (第1巡，牌墙余69张)。
+- 状态: 当前排名 1/4 (与一位差 0)。
+..."""
 
 # Inference
 inputs = tokenizer(input_text, return_tensors="pt")
@@ -184,10 +190,10 @@ A total of `192000` samples were used, with no general instruction data or self-
 
 Inference parameters: Temperature=0.1, Top_P=0.1
 
-**Metrics Explanation:**
+**Metrics explanation**:
 - Score: Max 500 points (1 point per correct sample, 0 for incorrect)
-- Full-match Rate: Samples where all 3 tests matched the dataset
-- Zero-score Rate: Samples where all 3 tests disagreed with the dataset
+- Full-match rate: Samples where all 3 tests matched the dataset
+- Zero-score rate: Samples where all 3 tests disagreed with the dataset
 
 #### Tile-Efficiency Test
 
@@ -238,7 +244,7 @@ Inference parameters: Temperature=0.6, Top_P=0.95
 
 ## License
 
-This model follows the Apache License 2.0.
+This model is licensed under Apache License 2.0.
 
 The training data comes from `pjura/mahjong_board_states`, which is licensed under `CC BY 4.0`. Please preserve the required attribution and citation when using it.