Create README.md

README.md

@@ -0,0 +1,40 @@
+---
+license: mit
+task_categories:
+- text-generation
+- feature-extraction
+language:
+- en
+tags:
+- gpt-oss
+- mla
+- calibration
+- fineweb-edu
+size_categories:
+- 10K<n<100K
+---
+
+# GPT-OSS Calibration Data (SWA/Full Splits)
+
+This dataset contains high-quality calibration artifacts for converting **GPT-OSS-20B** (and larger MoE models) from GQA to **Multi-Head Latent Attention (MLA)**.
+
+It was generated using **GTE-Large-v1.5** (8192 context window) on an L40S GPU to ensure long-range dependencies are preserved, addressing the "V_latent compression" quality gap.
+
+## Files
+
+| File | Description | Target Layers |
+| :--- | :--- | :--- |
+| **`calib_swa_golden.jsonl`** | Dense local coherence data (< 2048 tokens). | Sliding Window Attention Layers |
+| **`calib_full_golden.jsonl`** | Long-range dependency outliers (> 4096 tokens). | Full Attention Layers |
+| **`calib_mixed_golden.jsonl`** | A diverse mix of both (General purpose). | Fallback / Baseline |
+
+## Methodology
+1.  **Source:** [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) (SOTA reasoning/academic text).
+2.  **Embeddings:** `Alibaba-NLP/gte-large-en-v1.5` (8k context) used to encode documents.
+3.  **Filtering:** 
+    *   **Clustering:** MiniBatchKMeans to find representative centroids.
+    *   **Hard Negatives:** Top 5% outliers (highest distance from centroids) selected to maximize gradient signal during distillation.
+    *   **Splitting:** Data strictly stratified by length to match GPT-OSS SWA/Full layer requirements.
+
+## Usage
+Intended for Knowledge Distillation (Teacher -> Student) to repair Perplexity degradation during MLA conversion.