Add HuggingFace Hub checkpoint persistence - upload and download checkpoints between jobs

Add checkpoint resumption - automatically resume from latest checkpoint

Fix ReasoningStep attribute - use description not content

Fix SPECIAL_TOKENS usage - import at module level and use string literals

Fix ReasoningChain dataclass - add image field and defaults, fix collate function

Fix None image handling in collate function

Convert learning_rate to float explicitly

Remove dead code with direct config access

Add getattr defaults for all config parameters

Fix FineVision dataset loading with subset parameter

Implement full SFT, PRM, and RL training with dataset loading

Make train_prm and train_rl placeholders - dataset loading needs HF integration

Add **kwargs to train_prm and train_rl to accept config parameters

Fix train_prm and train_rl signatures to accept max_steps and learning_rate

Add placeholder train_sft method

Add type conversion for RLConfig parameters

Add type conversion and debug logging for config values

Fix gradient checkpointing for VLM models

Fix tokenize() calls to use actual tokenizer

Fix len() calls to use actual_tokenizer instead of processor

Fix tokenizer access for Processor objects

Add VLM support to trainer with auto-detection

Add training scripts and configs