Commit History

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

    renpas22
    3b46388
  2. Restore SFT to 10000 steps - will resume from checkpoint at 7000

    renpas22
    e9e0301
  3. Add checkpoint resumption - automatically resume from latest checkpoint

    renpas22
    5419afd
  4. Reduce training steps to fit within job timeout (SFT: 5k, PRM: 2k, RL: 3k)

    renpas22
    5cfd8d6
  5. Fix ReasoningStep attribute - use description not content

    renpas22
    f941008
  6. Fix SPECIAL_TOKENS usage - import at module level and use string literals

    renpas22
    464ac9b
  7. Fix ReasoningChain dataclass - add image field and defaults, fix collate function

    renpas22
    3024a91
  8. Fix None image handling in collate function

    renpas22
    d29f3e7
  9. Convert learning_rate to float explicitly

    renpas22
    9e7779a
  10. Remove dead code with direct config access

    renpas22
    ccd696b
  11. Add getattr defaults for all config parameters

    renpas22
    cd76323
  12. Fix FineVision dataset loading with subset parameter

    renpas22
    e47ae2c
  13. Implement full SFT, PRM, and RL training with dataset loading

    renpas22
    84a183c
  14. Make train_prm and train_rl placeholders - dataset loading needs HF integration

    renpas22
    7bff7cb
  15. Force fresh repository download with cache clearing

    renpas22
    714d05d
  16. Add **kwargs to train_prm and train_rl to accept config parameters

    renpas22
    917e40e
  17. Force download latest revision from HF repo

    renpas22
    f8fc68a
  18. Fix train_prm and train_rl signatures to accept max_steps and learning_rate

    renpas22
    41bcc92
  19. Add placeholder train_sft method

    renpas22
    85ab8c2
  20. Fix quantized model device handling in inference_scaling

    renpas22
    a745b26
  21. Add type conversion for RLConfig parameters

    renpas22
    f15c2d7
  22. Add type conversion and debug logging for config values

    renpas22
    c74a578

  1. Quote mixed_precision value

    renpas22
    37e8f2f
  2. Add inference config parameters

    renpas22
    2d1ba1a
  3. Skip .to(device) for quantized models with device_map

    renpas22
    fa9e543
  4. Add missing RL/PPO config parameters

    renpas22
    5af9eca
  5. Fix gradient checkpointing for VLM models

    renpas22
    0326431
  6. Fix tokenize() calls to use actual tokenizer

    renpas22
    8268436
  7. Fix working directory for HF Jobs environment

    renpas22
    bb7ed44

  1. Fix len() calls to use actual_tokenizer instead of processor

    renpas22
    4605c1b
  2. Fix tokenizer access for Processor objects

    renpas22
    b8bd3e8
  3. Add VLM support to trainer with auto-detection

    renpas22
    e20135f
  4. Fix model name to use valid Qwen2-VL model

    renpas22
    61dbc34
  5. Enable 8-bit quantization and reduce batch size for memory

    renpas22
    d4b8544
  6. Add required top-level config keys for trainer

    renpas22
    9770aa1
  7. Fix config path resolution for trainer initialization

    renpas22
    ec4cb07
  8. Fix function call arguments and trainer initialization

    renpas22
    879210b
  9. Fix config path resolution for HF Jobs

    renpas22
    82de57b
  10. Add utils directory

    renpas22
    da76488
  11. Auto-download repository in HF Jobs environment

    renpas22
    487225b
  12. Fix imports to work with sys.path

    renpas22
    83e0535
  13. Fix Python path for imports

    renpas22
    27b06f0
  14. Add training scripts and configs

    renpas22
    2b8876a
  15. initial commit

    Verified signature
    cd336ff