---
license: mit
datasets:
- K-and-K/knights-and-knaves
language:
- en
base_model:
- Xkev/gemma-3-1b-it-kk
pipeline_tag: text-generation
library_name: transformers
tags:
  - knights-and-knaves
  - gemma3
  - rl
---


# Gemma-3-1B-IT — Knights-and-Knaves BES

Paper Link: [https://arxiv.org/abs/2605.28814](https://arxiv.org/abs/2605.28814)

Post-trained model on top of [`Xkev/gemma-3-1b-it-kk`](https://huggingface.co/Xkev/gemma-3-1b-it-kk) using Bidirectional Evolutionary Search (BES) on the [Knights-and-Knaves](https://huggingface.co/datasets/K-and-K/knights-and-knaves) (K&K) logic-puzzle dataset.

For the SFT cold-start this was initialized from, see [`Xkev/gemma-3-1b-it-kk`](https://huggingface.co/Xkev/gemma-3-1b-it-kk).

## Training

- **Base model**: `Xkev/gemma-3-1b-it-kk`
- **Dataset**: K&K 5k train split
- **Framework**: [verl](https://github.com/volcengine/verl) `main_ppo` with a bidirectional goal-tree search agent loop
- **Search**: budget=200 rollouts, decompose interval=10, backward model `google/gemma-3-1b-it`
- **Hyperparameters**: `lr=1e-6`, batch=32, `ppo_epochs=1`, `clip_ratio=0.2`, `grad_clip=0.3`, `kl_coef=0`, `dtype=bf16`

## Intended use

Research on logical reasoning and post-training. Not intended for general dialog or production.

## License

MIT. Base model `google/gemma-3-1b-it` is governed by Google's [Gemma Terms of Use](https://ai.google.dev/gemma/terms), which still apply transitively to this model.