m0no1/dnd-35-mechanics-170k

D&D 3.5 Mechanics Dataset - FINAL VERSION

Ready for Llama 3.1 8B Training

This directory contains the final, balanced, deduplicated dataset for training Llama 3.1 8B to properly handle D&D 3.5 mechanics.

Dataset Statistics

Files

• final_train.jsonl - 170,018 training examples (42MB)
• final_val.jsonl - 3,000 validation examples (668KB)
• llama3_mechanics_final.yaml - Axolotl training config

Composition

• Total Examples: 173,018 (170k train + 3k val)
• Source:
  • 121,881 from o3's cleaned/audited data
  • 48,137 newly generated balanced examples
  • 3,000 validation (stratified by o3)

Distribution (Well-Balanced)

• Action (combat/skills/spells): 50.2%
• Social (dialogue/NPC): 32.8%
• Creative (narrative roleplay): 15.9%
• Clarification: 0.6%
• Other: 0.4%

Quality Metrics

• ✅ Zero duplicates (MinHash-LSH deduped at 0.85 threshold)
• ✅ All outputs valid JSON under 160 chars
• ✅ Natural language variety (no robotic patterns)
• ✅ Covers edge cases (long narration, ambiguous input)
• ✅ No placeholders or invalid data

Training Instructions

1. Copy dataset to training environment:

cp -r DND_MECHANICS_FINAL_DATASET_170K /workspace/data/

2. Install Axolotl if needed:

pip install axolotl

3. Validate config:

axolotl validate-config /workspace/data/DND_MECHANICS_FINAL_DATASET_170K/llama3_mechanics_final.yaml

4. Train:

accelerate launch -m axolotl.cli.train /workspace/data/DND_MECHANICS_FINAL_DATASET_170K/llama3_mechanics_final.yaml

Expected Results

With this balanced dataset, Llama 3.1 8B should:

• Correctly identify actions requiring dice rolls vs pure roleplay
• Handle natural language with contractions, hesitations, and variety
• Process long narrative descriptions without forcing mechanics
• Recognize ambiguous input and request clarification
• Output clean JSON for game mechanics integration

Important Notes

1. This dataset fixes the critical issue where Qwen3 14B incorrectly requires dice rolls for roleplay actions
2. The validation set is pre-stratified - do NOT use val_set_size in Axolotl
3. Monitor exact_match_json metric during training (target ≥0.9)
4. Dataset has been extensively audited by o3 for quality

Dataset Creation Process

1. Started with o3's 26k curated examples
2. Generated 125k+ examples with GPT-4o-mini
3. Applied MinHash-LSH deduplication (removed 120k+ duplicates)
4. o3 audited and cleaned (removed 27k invalid rows)
5. Generated 48k balanced examples to fix distribution
6. Final deduplication and merge

This dataset is the result of extensive iteration and quality control.