JohnGenetica/agentquery-gemma21b-reap-grpo-lora

AgentQuery FunctionGemma 270M Lab

This repository is the Hugging Face publication surface for the AgentQuery FunctionGemma 270M lab track.

Current contents:

• benchmark and holdout evidence from the AgentQuery repo
• training/evaluation manifests
• model card scaffolding for a future adapter or merged release

Current status:

• No weights are published in this revision.
• This repo exists to anchor the public/private release path and keep benchmark artifacts versioned on the Hub.

Source repository:

• server/reports/model_benchmark_matrix_gemma_holdout_baseline.json
• server/reports/sota_text2sql_training_set.jsonl
• server/reports/mechanistic_reap_report.json

Intended next step:

1. Train or adapt FunctionGemma locally, MLX, or HF Jobs.
2. Upload adapter or merged weights into this same repo.
3. Re-run holdout and function-calling smoke lanes against the published revision.