Tags: OpenPipe/ART
Tags
WIP: SFT (local backend) (#530) * SFT data iterator * Add SFT LR utils * train_sft skeleton * SFT Shape 0.1 * Add shuffle to SFTConfig * change SFT args order * Refactor SFT to accept batched trajectories Move batching and shuffling logic from SFTConfig into iterator functions. train_sft now accepts Iterable[List[Trajectory]] instead of individual trajectories, simplifying the API and making batch management more explicit. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com> * Tokenize SFT Batch * Add num_trainable_tokens to SFTBatch * draft train_sft * Flatten trajectory for train_sft * Tokenize SFT Batches support flat list and add padding * Fix max_length duplicate name issue * Remove unused file * remove unused typing * sft iterator * SFT Iterator * Use Unsloth for train on response * refactoring * implement local backend SFT training * Add SFT to Local Backend * avg loss * refactor, sft works good * remove logging * move tokenizer, update backend * update lr schedule and tests * refactor sft training from file * change batch sft * refactor step count based on checkpoints * update sft warmup script * fix model registration * make local random * refactor backend * refactor * update example * Pyright fix * remove iterate file epochs, refactor * refactor * add serverless endpoint * Rename training_folder_url to training_data_url * update defaults, change reporting * update lables * make sft to produce only one checkpoint step * refactor train from file * refactor * Refactor SFTTrainConfig * refactor * correctly register lora, fix unsloth proxy check * add sft train from file streaming * add openpipe qwen back * lint fix * calculate pbar * rename to training_data_url * accept model run_id from server * update optimizer hparams * add claude command * remove queue, add skills * add docs and colab example * move zero_grad * add final step arg * update docs * update docs and trajectories * lint fix * add cli skills * add chunking * lint fix * remove inline trajectories from skills * update chunking * change default chunk to 10 * remove leftovers --------- Co-authored-by: Angky William <awilliam@coreweave.com> Co-authored-by: Claude <noreply@anthropic.com>
PreviousNext