8Planetterraforming/Parameter-Golf-V7-Resolution-Layout-Compression-Reasoning

---
language:
- en
license: mit
tags:
- parameter-golf
- auxiliary-training
- compression-reasoning
- image-resolution
- layout-filtering
- multimodal-consistency
- synthetic-data
task_categories:
- text-generation
- question-answering
size_categories:
- 1K<n<10K
pretty_name: Parameter-Golf-V7-Resolution-Layout-Compression-Reasoning
---

# Parameter-Golf-V7-Resolution-Layout-Compression-Reasoning

## Overview

This is a synthetic auxiliary dataset for OpenAI Parameter Golf style experiments. It is **not** a replacement for FineWeb. It is designed to test whether a tiny auxiliary mix can reduce wasted next-token probability around compression reasoning, web-page signal filtering, display-vs-payload confusion, and multimodal preview planning.

The central lesson is deliberately narrow:

> Smaller rendered objects can reduce image/video payload, but smaller font size does not reduce the underlying text tokens. For language-model BPB, the useful target is shorter cleaner content, less repeated boilerplate, better tokenizer fit, and fewer wrong continuations.

## Why V7 exists

V6/B7 focuses on privacy filtering and web signal extraction. V7 extends that idea into resolution/layout/compression reasoning:

- distinguish **presentation** from **payload**,
- distinguish **preview filtering** from **final generation**,
- preserve exact discrete labels in cube/alphabet/optics-style multimodal tasks,
- avoid overclaiming from tiny file-size differences,
- keep Parameter Golf aligned with FineWeb BPB.

## Relation to Parameter Golf

Leaderboard performance is evaluated on FineWeb validation BPB. Therefore V7 should only be used as a small auxiliary probe:

- start: 0.25% or 0.5% V7 mix,
- conservative test: 1.0% V7 mix,
- only try 2–3% if seed42 improves and 3-seed mean confirms it,
- never replace FineWeb with V7.

## Schema

Each JSONL row has:

- `id`
- `task`
- `subcategory`
- `input`
- `target`
- `source_theme`
- `difficulty`
- `parameter_golf_role`
- `recommended_mix`

## Splits

See `stats.json` for exact counts and hashes.

## File layout

```text
data/train.jsonl
data/validation.jsonl
data/test.jsonl
stats.json
source_sanitization.md
upload_to_hf.md
convert_to_chat_format.py
parameter_golf_v7_probe_plan.md
```

## Important caveat

This dataset does **not** train image generation directly. It trains text behavior around planning, compression accounting, and signal extraction. Any improvement for image/video/3D workflows would be indirect through better instruction following and fewer wasteful generations.

# Parameter-Golf-V7-Bonus-Extended-Compression-Examples

Additional bonus dataset for Parameter Golf V7.

This file extends the V7 concept of:
- resolution-aware filtering,
- layout vs payload reasoning,
- DOCX/font-size observations,
- image preview budgeting,
- FineWeb-style noise removal.

---

## Technical Note (Important)

Font size can affect rendered documents (e.g. DOCX) due to style and layout metadata.
However, it does **not directly reduce language-model token count**.

For Parameter Golf, the correct interpretation is:

- remove repeated templates, styling markup, tracking strings, and low-value page chrome;
- use low-resolution previews only as a filtering stage;
- preserve semantic payload and exact symbolic constraints;
- keep FineWeb as the primary corpus;
- use this dataset strictly as a small auxiliary probe.

---

## Recommended Mixing Ratios

Start conservatively:

- 99.75% main corpus / 0.25% V7 bonus
- 99.50% main corpus / 0.50% V7 bonus
- 99.00% main corpus / 1.00% V7 bonus

Reject any mixture that degrades FineWeb seed BPB.

---

## Contents

- `v7_bonus_extended_examples.jsonl` — 930 examples
- `v7_bonus_stats.json` — counts and checksum
- `v7_bonus_probe_plan.md` — minimal evaluation plan

---

## Hugging Face Title

`Parameter-Golf-V7-Bonus-Extended-Compression-Examples`

---

## Short Description

V7 bonus dataset for Parameter Golf: resolution preview filtering, DOCX/font-size layout reasoning, payload vs presentation separation, FineWeb noise filtering, and BPB-safe compression heuristics.