[Bounty: 50 RTC] Training data generator for SFT pipeline

[Scottcjn]

Description

Port the SophiaCore data generation pattern to ShaprAI for supervised fine-tuning (SFT) training data generation.

Requirements

• Port the proven pattern from sophiacore_data_generator.py to work with ShaprAI's template system
• Generate ChatML-formatted training data with proper <|im_start|> / <|im_end|> tokens
• Support identity-weighted examples (personality-defining responses weighted higher in training)
• Customizable personality templates — users define their agent's voice, values, and behavioral boundaries

Acceptance Criteria

• shaprai/training/sft_generator.py module created
• Generates valid ChatML JSONL output
• Identity-weighted sampling: core personality examples appear 3-5x more frequently
• Template-driven: personality defined via YAML/JSON config, not hardcoded
• CLI command: shaprai generate-sft --template my_agent.yaml --output train.jsonl --count 1000
• Includes at least 3 example personality templates
• Compatible with HuggingFace TRL SFTTrainer format
• Unit tests for generator logic

Bounty

50 RTC — Paid on merge to main.

How to Claim

Comment on this issue to claim it. Submit a PR referencing this issue.

[sarahdack854]

/attempt #2

I would like to work on this bounty. I have experience with Python data generation, template-driven systems, CLI tools, and training data pipelines.

Plan:

1. Add shaprai/training/sft_generator.py
2. Generate valid ChatML JSONL output from YAML/JSON personality templates
3. Implement identity-weighted sampling so core personality examples recur 3-5x more often
4. Add CLI command: shaprai generate-sft --template my_agent.yaml --output train.jsonl --count 1000
5. Include at least 3 example personality templates
6. Add unit tests for generator logic and output validation

I will implement this now following the same repo workflow used successfully for #8 and #5.

[BinShao1988]

🎯 Bounty Claim: Issue #2

I have completed the SFT training data generator implementation!

✅ Delivered

• shaprai/training/sft_generator.py - Full SFT generator with:

  • ChatML format output (<|im_start|> / <|im_end|>)
  • Identity-weighted examples (3-5x higher weight)
  • Template-driven personality configuration
  • 3 example personality templates
  • CLI command: shaprai generate-sft
  • HuggingFace TRL SFTTrainer compatibility

• tests/test_sft_generator.py - Comprehensive unit tests

📋 PR

See PR #59 for the complete implementation.

Claiming: 50 RTC