-
Add HuggingFace Hub checkpoint persistence - upload and download checkpoints between jobs
renpas22
-
Restore SFT to 10000 steps - will resume from checkpoint at 7000
renpas22
-
Add checkpoint resumption - automatically resume from latest checkpoint
renpas22
-
Reduce training steps to fit within job timeout (SFT: 5k, PRM: 2k, RL: 3k)
renpas22
-
Fix ReasoningStep attribute - use description not content
renpas22
-
Fix SPECIAL_TOKENS usage - import at module level and use string literals
renpas22
-
Fix ReasoningChain dataclass - add image field and defaults, fix collate function
renpas22
-
Fix None image handling in collate function
renpas22
-
Convert learning_rate to float explicitly
renpas22
-
Remove dead code with direct config access
renpas22
-
Add getattr defaults for all config parameters
renpas22
-
Fix FineVision dataset loading with subset parameter
renpas22
-
Implement full SFT, PRM, and RL training with dataset loading
renpas22
-
Make train_prm and train_rl placeholders - dataset loading needs HF integration
renpas22
-
Force fresh repository download with cache clearing
renpas22
-
Add **kwargs to train_prm and train_rl to accept config parameters
renpas22
-
Force download latest revision from HF repo
renpas22
-
Fix train_prm and train_rl signatures to accept max_steps and learning_rate
renpas22
-
Add placeholder train_sft method
renpas22
-
Fix quantized model device handling in inference_scaling
renpas22
-
Add type conversion for RLConfig parameters
renpas22
-
Add type conversion and debug logging for config values
renpas22
-
Quote mixed_precision value
renpas22
-
Add inference config parameters
renpas22
-
Skip .to(device) for quantized models with device_map
renpas22
-
Add missing RL/PPO config parameters
renpas22
-
Fix gradient checkpointing for VLM models
renpas22
-
Fix tokenize() calls to use actual tokenizer
renpas22
-
Fix working directory for HF Jobs environment
renpas22
-
Fix len() calls to use actual_tokenizer instead of processor
renpas22
-
Fix tokenizer access for Processor objects
renpas22
-
Add VLM support to trainer with auto-detection
renpas22
-
Fix model name to use valid Qwen2-VL model
renpas22
-
Enable 8-bit quantization and reduce batch size for memory
renpas22
-
Add required top-level config keys for trainer
renpas22
-
Fix config path resolution for trainer initialization
renpas22
-
Fix function call arguments and trainer initialization
renpas22
-
Fix config path resolution for HF Jobs
renpas22
-
Add utils directory
renpas22
-
Auto-download repository in HF Jobs environment
renpas22
-
Fix imports to work with sys.path
renpas22
-
Fix Python path for imports
renpas22
-
Add training scripts and configs
renpas22
-
initial commit