CoolFace
Datasetpublic

8Planetterraforming/Parameter-Golf-V19-Family-Training-Mix

Parameter-Golf V19 — Family Training Mix V19 is an English training micro-mix package built from the visible 8Planetterraforming dataset family for OpenAI Parameter Golf work. This package is not just a control note. It contains real training files: data/train.jsonl data/validation.jsonl data/test.jsonl data/plain_text/train.txt data/plain_text/v19_micro_0p02pct.txt data/plain_text/v19_micro_0p05pct.txt data/plain_text/v19_micro_0p10pct.txt Core objective The… See the full description on the dataset page: https://huggingface.co/datasets/8Planetterraforming/Parameter-Golf-V19-Family-Training-Mix.

sourceHugging Faceupdated 5mo agoView on Hugging Face
0likes50downloads
Dataset Card

Parameter-Golf V19 — Family Training Mix

V19 is an English training micro-mix package built from the visible 8Planetterraforming dataset family for OpenAI Parameter Golf work.

This package is not just a control note. It contains real training files:

  • —data/train.jsonl
  • —data/validation.jsonl
  • —data/test.jsonl
  • —data/plain_text/train.txt
  • —data/plain_text/v19_micro_0p02pct.txt
  • —data/plain_text/v19_micro_0p05pct.txt
  • —data/plain_text/v19_micro_0p10pct.txt

Core objective

The objective is to compress the most useful lessons from all 26 user datasets into a small English corpus that can be mixed into Parameter Golf training experiments.

The target remains a legal lower val_bpb on official FineWeb validation. V19 should never be used as validation data or hidden model storage.

Current leaderboard target gap

The package uses the current public leaderboard pattern where the top listed score is 1.0611 BPB. The target 0.9000 BPB is 0.1611 BPB below that frontier.

That gap is too large for dataset text alone. V19 is designed to help a stronger stack, not replace it.

Strongest V19 training path

  1. 1.Start from the current SOTA family: BOS fix, SmearGate, LQER, SparseAttnGate/PolarNS/FusedCE, CaseOps, SP8192, phased legal TTT, recurrence, parallel residuals, and compression-aware quantization.
  2. 2.Add V19 only as a tiny micro-mix.
  3. 3.Use the V13 idea to annotate official FineWeb training shards with structural metadata.
  4. 4.Prioritize V8/V9/V10 web-signal micro-mix records because they are closest to FineWeb-like text.
  5. 5.Keep V16/V17/V18 records as training text that teaches legal gates, experiment routing, and failure prevention.

Files

  • —data/dataset_index.jsonl: 26 numbered source datasets with short description, key content, BPB hypothesis, mix rate, and risk gate.
  • —data/leaderboard_comparison.jsonl: current public leaderboard comparison and gap to 0.9000 BPB.
  • —data/priority_actions.jsonl: the main research actions for the next run.
  • —docs/EXTENDED_DESCRIPTION_EN.md: long English explanation.
  • —docs/MIXING_PLAN_EN.md: practical mixing plan.
  • —scripts/validate_v19_dataset.py: schema validator.
  • —scripts/preflight_v19_run.sh: H100 preflight gate.

Legal use boundary

V19 is training auxiliary data. It is not validation. It is not a replacement for the official FineWeb validation split. Any leaderboard claim still needs official scoring, full logs, artifact-size proof, and statistical evidence.