8Planetterraforming/solutions-training-v2

Parameter Golf Auxiliary Dataset — PG-CCE-200

PG-CCE-200 is a compact auxiliary training dataset designed to target three failure modes that hurt next-token modeling quality in compact language models:

1. Hallucination / overconfident guessing
2. Weak long-context state tracking
3. Fragility on exact discrete structure (numbers, filenames, delimiters, ordered lists, exact project state)

This dataset was derived from real failure patterns observed during iterative model training, repository management, RunPod debugging, and exact arithmetic / structured reasoning tasks.

It is intended as a helper dataset, not a replacement for the main competition corpus.

Main idea

The dataset teaches the model to prefer:

• clarification before guessing
• separation of verified vs unverified information
• canonical project-state tracking
• short, high-signal stepwise guidance
• exact preservation of filenames, delimiters, lists, units, and numeric transformations

These behaviors can reduce wasted probability mass on plausible-but-wrong continuations and improve stability on structured text.

Recommended use

Use this dataset only as a small auxiliary mixture with the main training corpus.

Recommended starting ratio:

• 97% main corpus
• 3% PG-CCE-200

Safer exploratory ratios:

• 99% / 1%
• 98% / 2%

Do not replace the main corpus with this dataset.

Splits

• Train: 160 examples
• Validation: 20 examples
• Test: 20 examples

Categories

1. uncertainty_calibration

Examples teach the model to:

• ask for missing high-impact variables
• separate confirmed facts from guesses
• avoid pretending to know when the prompt is underspecified

2. context_state_management

Examples teach the model to:

• use the latest verified project state
• prefer canonical state over stale fragments
• answer in short, stepwise fashion when the user is executing commands
• avoid branching into many future instructions before the current blocker is resolved

3. exact_discrete_reasoning

Examples teach the model to:

• preserve exact filenames and command tokens
• count exactly
• preserve ordering and delimiters
• compute exact cube-volume scaling by powers of two
• keep project-state comparisons exact

Fields

Each record contains:

• id
• category
• task
• instruction
• input
• target
• rationale
• difficulty
• language

Example record

{
  "id": "pgcce_0001",
  "category": "uncertainty_calibration",
  "task": "clarify_before_recommendation",
  "instruction": "Answer with calibrated uncertainty. Ask for the missing variables before recommending.",
  "input": "User: Which shirt color will suit blue eyes best?",
  "target": "Clarify first: What is your skin tone, hair color, trouser color, and occasion? Eye color alone is not enough to choose the best shirt.",
  "rationale": "The question is underspecified. Ask about missing variables before recommending.",
  "difficulty": "easy",
  "language": "en"
}

Why this can help BPB

This dataset is based on the idea that compact models often waste probability mass on:

• generic guesses
• stale context references
• broken exact strings
• wrong numbers that are close in style but wrong in content

By teaching the model to be more selective, more state-aware, and more exact on structured text, the dataset may reduce entropy on downstream outputs.

Limits

This dataset alone will not magically produce state-of-the-art BPB.
It is a supporting dataset for training discipline and structured behavior.

It is most useful when:

• the main pipeline already works
• the main corpus remains dominant
• the user wants a small, high-signal auxiliary component

Included files

• train.jsonl
• validation.jsonl
• test.jsonl
• schema.json
• dataset_info.json
• prepare_aux_text.py
• RUNPOD_NOTES.md

Suggested next step

Convert the JSONL examples into a compact text format and mix them conservatively into the main training corpus.