---
license: apache-2.0
base_model: convaiinnovations/laya-multilingual
datasets: [paintedwolfcode/bialy-dataset]
tags: [coding-agents, tool-selection]
---

# Bialy decision heads open1-b5-b7g-e4

Tuned heads built on [Laya](https://github.com/NandhaKishorM/laya) by Convai Innovations,
over its frozen multilingual encoder that Painted Wolf Code
asks as it works: which loadable tool schemas a turn will need, which
instruction units it can leave out, and what kind of work it is (`turn-load`);
how relevant each skill and tool card is to a request (`unit-rank`); and which
code units best answer a request, blended into the text-match order of code
summaries, repository maps, and project search (`code-rank`). The engine
(`bialy`) loads them beside the backbone; a head trained over another
backbone is refused.

Trained on [paintedwolfcode/bialy-dataset](https://huggingface.co/datasets/paintedwolfcode/bialy-dataset) open1-b7g-e4: `turn-load`
and `unit-rank` on its session rows, labeled by what open-weights models did
in coding-agent sessions and scored by an open-weights judge; `code-rank` on its
code-rank pairs, requests open-weights models wrote for code units in the same
repositories. Trained and replayed with Painted Wolf Code at commit
68e19e34e7b85a459a2a007af52f3611d6996089; packaged with https://github.com/paintedwolf-ai/bialy.

## Heads

- `code-rank.safetensors`: open1-code-rank, over jhu-clsp/mmBERT-base, sha256 `abe235c347a223e957a0f8b7b979e3a584aebab76ee239e5f62117e839b9c40a`
- `guide-load.safetensors`: open1-turn-load-B7G-release-independent, over jhu-clsp/mmBERT-base, sha256 `e0305aa8b62ddc144774785f2e0d3710604af65cd21015662428870f8d34b76e`
- `turn-load.safetensors`: open1-turn-load-B5-release-independent, over jhu-clsp/mmBERT-base, sha256 `550f30b94d5c6a8680d2cb2a607ebc9ee33e709c7679f72606d1e9fda17f2e90`
- `unit-rank.safetensors`: open1-unit-rank-e4-dense1, over jhu-clsp/mmBERT-base, sha256 `c62674993eabba94660eccfafbb6e04d1549adc7335656c0ddefa7f068fa58f7`

## Results

Replayed through the shipped engine on sets neither head trained on. Tool
columns are precision / recall / F1 of the loadable tools a turn used, at
the catalog's load threshold; guide omission precision is the share of
omitted instruction units the turn did not need; need MRR ranks the tools a
`request_tools` need went on to use.

| Set | Heads | tools P/R/F1 | macro R | loads/turn | guide omit P | kind acc | need MRR |
|---|---|---|---|---|---|---|---|

`code-rank`, on repositories it never trained on: the rank of the code unit a
request was written for, in the site's text-match order and blended with the
head.

| Report | Pairs | MRR text / blended | hit@1 text / blended | improved / regressed |
|---|---|---|---|---|

## Check it yourself

Each row of the table comes from a replay report shipped in `eval/`.
`bialy audit heads` (from https://github.com/paintedwolf-ai/bialy) replays these heads through the
Painted Wolf Code engine on a CPU over the dataset's held-out split and
compares every metric with the shipped report.