feat(skills): Create training-methods skill documenting 8 stable training methods

[akaszubski]

Summary

Create a new training-methods skill that documents all 8 stable training methods with their data requirements, use cases, and recommended training order.

Context

ReAlign supports 8 stable training methods but there's no centralized skill to help users choose the right method for their goal.

What Does NOT Work

• No centralized skill documenting all training methods
• Users don't know which method to use for their goal
• Data format requirements scattered across code
• Training order not documented

Implementation Approach

Create .claude/skills/training-methods.md documenting:

8 Stable Training Methods

Method	Purpose	Data Format	Use Case
SFT/LoRA	Base capabilities	instruction + output	Knowledge injection
DPO	Preference alignment	chosen + rejected	Behavior change
ORPO	Preference (simple)	chosen + rejected	No reference model needed
GRPO	Verifiable rewards	prompt + scored responses	Math/code verification
CPO	Conservative preference	chosen + rejected	Reduce distribution shift
RLVR	Verified rewards	problem + solution + verify	Complex reasoning
Abliteration	Remove constraints	harmful + harmless	Uncensoring
Activation Steering	Runtime control	inference-time	Reversible, composable

Data Formats

# SFT
{"instruction": str, "output": str}

# DPO/ORPO/CPO
{"prompt": str, "chosen": str, "rejected": str}

# GRPO
{"prompt": str, "responses": [{"text": str, "score": float}]}

# RLVR
{"problem": str, "solution": str, "responses": [{"text": str, "correct": bool}]}

Recommended Training Order

1. SFT (base) → Instruction following
2. DPO (alignment) → Behavior change
3. GRPO/RLVR (verification) → Reasoning improvement
4. Calibration → Uncertainty handling
5. Anti-hallucination → Reduce confabulation

Backend Support Matrix

Method	MLX	PyTorch	Cloud
LoRA	✅	✅	✅
DPO	✅	✅	✅
ORPO	✅	✅	✅
GRPO	✅	✅	✅
CPO	✅	✅	✅
RLVR	✅	❌	❌
Abliteration	✅	✅	❌
Steering	✅	❌	❌

Acceptance Criteria

• All 8 stable methods documented
• Data format examples for each method
• Training order recommendation with rationale
• CLI commands for each method
• Backend support matrix

Related

• Issue feat(skills): Update data-curation-workflow skill with A-grade 9-stage pipeline #305 (data-curation-workflow)

[akaszubski]

Implemented in PR #324 (batch-20260205-233714)