Download README.md from VGGamerz/KevIN: direct link, hf CLI and curl.
- Browser
- Download file 2.59 kB
-
https://huggingface.co/VGGamerz/KevIN/resolve/main/README.md
- Command line
-
hf download hf://VGGamerz/KevIN/README.md
-
curl -L -o README.md https://huggingface.co/VGGamerz/KevIN/resolve/main/README.md
license: apache-2.0
base_model: jaredpalmer/kev-0.8b
tags:
- decision-model
- lora
- multiple-choice
- wikipedia
KevIN
KevIN is a fine-tuned Kev-0.8B decision model for choosing links during Wikispeedia navigation. Given a target, current page and candidate links, it scores the supplied options. It does not generate text or images. Its trained LoRA adapter and pointer head build on the frozen Qwen/Qwen3.5-0.8B-Base backbone.
Checkpoint and training
This release contains RL update 150 from run 20261023T200829.319603Z-rl, selected as the run's best checkpoint on a fixed 64-pair development sample. It starts from full-SNAP SFT update 2,000, which was initialized from released Kev 0.8B v1.0. SFT used processed human routes from the SNAP Wikispeedia dataset; RL used on-policy REINFORCE on training pairs. The full dataset preparation considered 51,318 completed routes, of which 50,407 passed route validation. This uses the dataset's 4,604-article graph; it is not a full-Wikipedia navigation model. The adapter and decision head are stored separately in adapter_model.safetensors and head.pt. The checkpoint uses temperature 1.0; its output probabilities have not been separately calibrated for this task.
Evaluation and limitations
The checkpoint solved 62/64 pairs (96.88%) in its development selection sample and 2,619/2,867 pairs (91.35%) on the expanded held-out test split, within a 12-click limit. Evaluation used the SNAP graph, 1,500-character page excerpts and a tournament over candidate links; it is not an evaluation on current Wikipedia. Pages and targets can appear across splits, so these figures do not establish generalization to unseen articles. Performance depends on the navigation environment and choice presentation; the development figure comes from a small selection sample.
Use
Install Kev at revision 6b719c3c3f367295f6ef336f4f751cf5ff970abc with its serving dependencies, then run:
python -m kev.serve --run VGGamerz/KevIN --port 8008
Kev loads the adapter and pointer head and downloads the base model at the revision recorded in the checkpoint. Use Kev's System One API to supply the state and candidate choices. Loading the LoRA with generic PEFT alone does not load the pointer head or provide Kev's decision API.
This repository is an inference export. It does not include optimizer state or the data and run artifacts needed to resume the original training run.