Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,39 @@
+---
+license: gemma
+base_model: google/gemma-4-E4B-it
+tags:
+- gguf
+- llama.cpp
+- agent
+- tool-use
+- agenda-parser
+---
+
+# agenda-parser-lite (GGUF)
+
+LoRA SFT fine-tune of [`google/gemma-4-E4B-it`](https://huggingface.co/google/gemma-4-E4B-it) for the
+**Agenda Parser** research agent's ReAct tool-calling protocol — the model emits a
+single JSON action `{"thought", "tool", "args"}` per step over a public-meeting
+agenda packet. Quantized to **Q8_0** for [llama.cpp](https://github.com/ggml-org/llama.cpp).
+
+## Training data
+
+Distilled from strong teacher models (Kimi k2.6, DeepSeek 4 pro via OpenCode Go) on
+the [`rdubwiley/agenda-parser-tool-traces`](https://huggingface.co/datasets/rdubwiley/agenda-parser-tool-traces) trace dataset
+(faithfulness-judged). One agent step per example; loss on the assistant JSON span.
+The source agenda packets the traces were generated over are published in that
+dataset's [`source_packets/`](https://huggingface.co/datasets/rdubwiley/agenda-parser-tool-traces/tree/main/source_packets)
+folder.
+
+## Run
+
+```bash
+huggingface-cli download rdubwiley/agenda-parser-lite agenda-parser-lite-Q8_0.gguf
+llama-server -m agenda-parser-lite-Q8_0.gguf --jinja     # --jinja enables the chat/tool template
+```
+
+## Intended use & limitations
+
+For driving the Agenda Parser agent (in-process llama.cpp backend) over uploaded
+agenda PDFs. Not a general assistant; answers should be checked against the source
+packet. Inherits the [Gemma Terms of Use](https://ai.google.dev/gemma/terms).