Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,57 @@
+---
+license: apache-2.0
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- hermes-agent
+- tool-selection
+- function-calling
+- lora
+- agent
+pretty_name: Hermes Tool Selector
+size_categories:
+- n<1K
+---
+
+# Hermes Tool Selector
+
+Dataset for fine-tuning LoRA adapters to improve agent tool selection. Extracted from real [Hermes Agent](https://github.com/NousResearch/hermes-agent) sessions.
+
+## Dataset
+
+| Split | Examples | Source |
+|---|---|---|
+| `train.jsonl` | 182 (augmented) | Real Hermes sessions + synthetic augmentation |
+| `train_raw.jsonl` | 59 (raw) | Real sessions, no augmentation |
+| `eval.jsonl` | 15 | Held-out benchmark cases |
+
+## Format
+
+Each line is a JSON object:
+
+```json
+{
+  "messages": [
+    {"role": "system", "content": "..."},
+    {"role": "user", "content": "..."},
+    {"role": "assistant", "content": "..."}
+  ]
+}
+```
+
+## Training
+
+Used with [Axolotl](https://github.com/axolotl-ai-cloud/axolotl) for QLoRA fine-tuning:
+
+```yaml
+base_model: Qwen/Qwen2.5-7B-Instruct
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+```
+
+## Auto-updating
+
+Collected automatically every 6 hours from live Hermes sessions.