8Planetterraforming/solutions-training-v3

Parameter Golf Auxiliary Dataset V3 (PG-CCE-V3)

This dataset is a stronger V3 auxiliary training set built primarily from the user's own error analyses and concrete examples.

Core training themes

The dataset expands three user-grounded failure modes:

1. 1.Hallucination / guessing instead of calibrated uncertainty
2. 2.ask for missing context before answering
3. 3.separate confirmed facts from inferred or missing facts
4. 4.avoid confident recommendations when critical variables are unknown

1. 1.Weak long-context state management
2. 2.keep one canonical project state
3. 3.prioritize the latest verified run over older chat fragments
4. 4.answer step-by-step instead of giving many dependent commands at once
5. 5.compress responses when many questions are asked together

1. 1.Fragility in exact, discrete reasoning
2. 2.exact numbers
3. 3.exact file names and paths
4. 4.exact command strings
5. 5.exact continuation of power-of-eight volume patterns
6. 6.exact arithmetic shortcuts based on decimal complements

Why this dataset exists

The goal is to reduce entropy in model behavior by teaching:

• —better uncertainty calibration
• —better project-state tracking
• —better exactness in commands, numbers, and structured outputs

This is intended as an auxiliary dataset for Parameter Golf style training. It is not a replacement for the main competition dataset.

Recommended use

Start with a small mixture:

• —97% main corpus
• —3% PG-CCE-V3

Increase only if validation shows a real benefit.

File structure

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

Record format

Each JSONL record contains:

• —id
• —family
• —source_theme
• —language
• —instruction
• —input
• —target
• —tags

Design note

This V3 version is intentionally grounded in concrete user examples:

• —clothing/color recommendations with missing variables
• —project-state drift between older and newer BPB values
• —RunPod command discipline
• —cube-capacity sequences growing by ×8
• —decimal-complement calculation shortcuts
• —exact file names and copy commands

That grounding is deliberate: the purpose is to train reliability on the exact classes of mistakes the user repeatedly observed.