Losa10/Minecraft-1.20.1-forge-modding

Forge-SLM Dataset v2

Minecraft Forge 1.20.1 Mod Development Training Dataset for Qwen3.5-4B (DeltaNet Hybrid)

Notes on Metrics

• Forge API Specificity is weighted down by bug_fix records (2.7/10) which are code fragments without full class context. New records score 4.0-6.0/10.
• Code Compliance for new records: 9.7/10, 100% pass rate
• Think-Code Coherence improved from 4.9 → 8.7 for new records through programmatic + LLM regeneration
• The 22% "zero API records" in the full dataset are from bug_fix type which intentionally contains small code snippets

Dataset Statistics

Template Type Distribution

Type	Count	Percentage
bug_fix	9,144	53.3%
extend	2,192	12.8%
feature_addition	1,949	11.4%
from_scratch	1,426	8.3%
migrate	810	4.7%
fix_bug_injected	741	4.3%
fix_bug_simple	522	3.0%
issue_fix	374	2.2%

Category Distribution (Top 10)

Category	Count
item	2,733
block_entity	2,046
block	1,798
screen	1,370
renderer	1,328
network	1,223
mixin	1,167
container	895
sound	689
data_gen	616

File Contents

Core Training Data

• forge_sft_train.json — 17,159 SFT training records
• forge_sft_val.json — 1,126 SFT validation records
• forge_grpo_train.json — 17,159 GRPO training prompts
• forge_grpo_val.json — 1,126 GRPO validation prompts

Enriched Subsets (kimi_k2.6 format)

• sft_train_enriched.json — 200 enriched SFT records
• sft_val_enriched.json — 20 enriched SFT val records
• grpo_train_enriched.json — 50 enriched GRPO records
• grpo_val_enriched.json — 20 enriched GRPO val records

Format Specification

SFT Format

{
  "conversations": [
    {"role": "user", "content": "Task description with code"},
    {"role": "assistant", "content": "<think/>Analysis of the code...\n</think/>\n\n<search>\noriginal code\n</search>\n\n<replace>\nmodified code\n</replace>"}
  ],
  "metadata": {
    "template_type": "extend|feature_addition|from_scratch|migrate|fix_bug_injected|fix_bug_simple|issue_fix|bug_fix",
    "category": "item|block|block_entity|screen|network|...",
    "lang": "java",
    "quality_score": 0.0-1.0,
    "estimated_tokens": 1234,
    "difficulty": "easy|medium|hard|expert",
    "response_format": "search_replace|code_block",
    "computed_metrics": {
      "forge_api_score": 5.0,
      "code_compliance": 9.7,
      "think_coherence": 8.7
    }
  }
}

GRPO Format

{
  "prompt": "Task description with code",
  "type": "extend",
  "category": "item",
  "estimated_tokens": 1234,
  "template_type": "extend",
  "lang": "java",
  "quality_score": 0.73,
  "difficulty": "medium"
}

Think Format

All assistant responses use the <think/>...</think/> format for chain-of-thought reasoning before code. The think block contains:

• Specific Forge API references used in the code
• Key implementation decisions
• Critical constraints or edge cases

Quality Assurance Pipeline

1. Source code from 4,919 real Forge 1.20.1 mod files (17 projects)
2. Static analysis: Forge API specificity, code compliance, think-code coherence
3. NeoForge contamination check (corrected patterns)
4. Deduplication (SHA-256 content hash)
5. TODO/FIXME artifact removal
6. Computed quality scores (not random)

Target Model

• Architecture: Qwen3.5-4B (DeltaNet Hybrid: 45 DeltaNet + 15 GQA layers)
• Think format: <think/>...\n</think/>
• Code format: Search-replace diffs (<search>/<replace>) or code blocks

Known Limitations

• bug_fix records (53%) are code fragments with lower API density — this is inherent to bug-fix snippets
• issue_fix category reduced to 374 records (from 372 in v6.1) — quality over quantity
• Some from_scratch records use real source code as the "answer" — this is by design (source-grounded)
• LLM-as-Judge pass rate not yet measured (API rate-limited during processing)