8Planetterraforming/parameter_golf_v13_fineweb_systemaware_core_control

Parameter Golf V13 — FineWeb System-Aware Core + Control

This package is an English-first V13 working kit for Parameter Golf. It is designed to be stronger than a pure lesson corpus while still staying honest about what it is and what it is not.

Status

This package does not claim a measured 0.81 BPB result.
It treats 0.81 BPB as a research target that still requires:

• a legal val_bpb computation,
• a reproducible run under the 16 MB artifact cap,
• training under the 10-minute budget on 8×H100,
• enough logs to support the official statistical threshold for a new record claim.

As of 2026-04-23, the public leaderboard in the repository shows 1.0810 for the top public record, and new SOTA submissions must beat the previous record by at least 0.005 nats with evidence at p < 0.01, unless the change is only a systems-speed optimization with no ML change. See source_notes/sources.md for the exact source links and dates.

Why V13 is structured this way

V11 and V12 are useful as preparation and synthesis layers, but they are still mostly meta-data in the broad sense: they train reasoning, interpretation, and discipline.
V13 is meant to be more practical:

1. Core path: annotate the official FineWeb training data with structure-aware metadata that can improve packing, curriculum, and analysis without pretending that annotation alone is a score.
2. Control path: keep a compact English auxiliary dataset that teaches the system how to talk about legality, cooling, throughput, TTT, tokenizer risks, ablations, and PR hygiene without overclaiming.

This package ships the control path as data and the core path as code.
That is deliberate. Without direct access to the official challenge text source during packaging, it would be dishonest to pretend the package already contains a final official-core corpus.

What is included

• data/v13_control_en.jsonl
  A compact English control dataset with 120 records. It is not meant to replace the official FineWeb train set. It is meant to keep experiment language, logging, and reasoning sharp.

• scripts/annotate_fineweb_v13.py
  A lightweight annotator for raw FineWeb-like text exports. It computes structural fields such as doc_family, markup_ratio, code_ratio, numeric_ratio, entropy_band, pack_group, and curriculum_bucket.

• scripts/validate_v13_dataset.py
  Checks schema consistency, missing fields, and duplicate IDs in the control dataset.

• scripts/record_gate_helper.py
  A small helper for thinking about record threshold math and seed-level evidence. It does not replace official review.

• meta/schema.yaml
  Field schema for the control dataset and for the annotator output.

• meta/package_manifest.yaml
  Human-readable package metadata.

• docs/EXTENDED_EXPLANATION_EN.md
  The operating logic behind V13 in plain English.

• docs/HF_UPLOAD_NOTES.md
  Notes for uploading to Hugging Face without breaking the dataset viewer again.

• templates/ablation_log_template.csv
  A practical log template for seed-level ablations.

• templates/pr_readme_skeleton.md
  A clean starting point for a future record or non-record PR.

Intended use

Use V13 in this order:

1. Replicate a strong current stack.
2. Keep the official training and validation setup fixed unless you are explicitly running a documented ablation.
3. Use the annotator to build metadata over the full train corpus, not to hide unexplained filtering.
4. Use the control dataset to train or fine-tune reasoning, documentation style, and audit behavior around:
   • score-first TTT,
   • exact val_bpb semantics,
   • throughput and thermal stability,
   • tokenizer and byte-accounting risks,
   • ablation discipline,
   • PR structure.

Not intended as

• proof of a new leaderboard record,
• a substitute for the official FineWeb train/validation benchmark,
• a justification for score-after-adapt evaluation,
• a permission slip for casual tokenizer edits,
• a claim that cooling directly changes cross-entropy.

Design principles

• Legal first
• Exact wording over hype
• Full-train annotation over hidden filtering
• Systems support is real, but not magic
• One variable per ablation
• Seed-level evidence before bold claims

Viewer safety note

This README uses explicit configs.data_files YAML so Hugging Face should point the viewer only at data/v13_control_en.jsonl.
Structured metadata files are stored as .yaml and .md instead of extra .json datasets to reduce schema collisions.

Recommended next step

Run the annotator on an official or verified FineWeb text export, inspect the family counts, and only then decide whether any curriculum or packing policy deserves a real 3-seed ablation.