model-specs/blbooksgenre-spec

---
license: mit
language:
- en
task_categories:
- text-classification
tags:
- hf-model-spec
- draft
# Linked artefacts
- dataset:biglam/blbooksgenre
- model:TheBritishLibrary/bl-books-genre
pretty_name: "Model spec — BL Books Genre classifier"

configs:
- config_name: samples
data_files:
- split: samples
path: data/samples.jsonl

spec_version: hf-model-spec-0.1
status: draft
task: text-classification
labels:
- Fiction
- Non-fiction
data:
training_repo: biglam/blbooksgenre
config: title_genre_classifiction
input_field: title
label_field: label
split: train
samples:
path: data/samples.jsonl
count: 20
has_labels: true
description: "10 fiction + 10 non-fiction titles, stratified random sample"
eval:
metric: accuracy
target_value: 0.94
comparison_baseline: TheBritishLibrary/bl-books-genre # ~0.94 accuracy on its own held-out
held_out_split_strategy: "construct 80/20 stratified on label from training data"
constraints:
max_params: 500_000_000
license: cc0-1.0
budget_usd: 1
output:
target_org: small-models-for-glam
model_name: bl-books-genre-classifier
card_required_fields:
- pipeline_tag
- id2label
- label2id
- model_index_with_accuracy_and_macro_f1
- confusion_matrix
- spec_repo_commit_sha
review_gates:
- "Before push_to_hub: surface held-out accuracy + confusion matrix for sign-off."
- "If accuracy < 0.90, pause and propose a different approach rather than retrying the same recipe."
---

# British Library 19th-century book genre classifier

> **🧪 Draft — `hf-model-spec-0.1`** · Personal working draft, not an HF convention. Feedback via Community tab.

## Purpose

A binary text classifier that takes a 19th-century book title and outputs Fiction or Non-fiction. Designed for re-classifying British Library catalogue metadata — especially titles where the existing Dewey or genre annotation is missing or unreliable. Victorian-era English-language publications, BL cataloguing conventions; titles typically 5–30 words.

## Data

`biglam/blbooksgenre` has two other configs (`annotated_raw`, 4,398 rows with `Can't tell` / `both fiction and non-fiction` labels; `raw`, 55,343 unfiltered records) — use only `title_genre_classifiction` (1,736 rows) for v1. The other configs become relevant for a v2 active-learning loop but are out of scope here.

Sample rows (20, stratified — 10 Fiction + 10 Non-fiction) live in [`data/samples.jsonl`](./data/samples.jsonl).