| # NOTICE | |
| Lean Laguna is a **serving recipe + benchmark harness + reusable RL environment** built on top of | |
| Poolside's released Laguna XS.2. It does **not** redistribute model weights. | |
| ## Built on | |
| - **Laguna XS.2** — `poolside/Laguna-XS.2` (Apache-2.0). The base model; not included here. | |
| - **DFlash speculator** — `poolside/Laguna-XS.2-speculator.dflash`. The 0.6B draft model used for | |
| speculative decoding; not included here. The Laguna DFlash *speculator checkpoint* was trained by Poolside. | |
| - **DFlash method** — the speculative-decoding drafting method, integrated in vLLM via | |
| `--speculative-config '{"method":"dflash", ...}'`. | |
| ## What is original here (Apache-2.0) | |
| The benchmark/serving harness (`scripts/`, `bench/`, `evals/`), the measured A/B results (`results/`), | |
| the `spec_rl` verifiers environment (`spec_rl/`), and the endpoint/configuration seam (`configs/`). | |
| Under greedy decoding the outputs are byte-identical to the base model — the speedup is lossless. | |
| ## Cite | |
| Built for the Poolside Research Hackathon. See `README.md` for the method, measured results, and | |
| reproduction steps. | |