SFT (local backend) #530
Open
SFT (local backend) #530
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
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>
src/art/types.py
Outdated
|
|
||
|
|
||
| class SFTConfig(pydantic.BaseModel): | ||
| learning_rate: float = 1e-4 |
Collaborator
There was a problem hiding this comment.
Remove custom_lr_schedule
Make learning_rate: float | list[float]
| Used to identify where assistant turns begin (train on responses only). | ||
| """ | ||
|
|
||
| instruction_part: str |
Collaborator
There was a problem hiding this comment.
We probably can keep this class as empty?
Unsure if instruction_part and response_part is a good fit for experimental feature
src/art/local/backend.py
Outdated
| batch_size = 2 # Default to 2 for SFT | ||
|
|
||
| # Determine learning rates | ||
| if config.custom_lr_schedule and len(config.custom_lr_schedule) > 0: |
Collaborator
There was a problem hiding this comment.
- Refactor/Remove
custom_lr_schedule.learning_rateisfloat | list[float] - Add validation for
num_learning_rate==num_batches
src/art/unsloth/service.py
Outdated
|
|
||
| # Save checkpoint after training | ||
| # Name checkpoint by final training step: starting_step + num_batches | ||
| final_step = get_step_from_dir(self.output_dir) + len(sft_batches) |
Collaborator
There was a problem hiding this comment.
Checkpoint step should be still incremented by 1.
Checkpoint step != Gradient step
| response_part="<|im_start|>assistant\n", | ||
| ), | ||
| # Qwen 3 models (with thinking tokens) | ||
| "Qwen/Qwen3-8B": ModelConfig( |
Collaborator
There was a problem hiding this comment.
- How we decide to support all of this model?
- Prefer to keep it simple and start with model that's widely use in OpenPipe Platform and ART?
- Research Qwen chat template, iirc
<think></think>only show up at the last turn. We may need to remove<think></think>inresponse_partin Qwen.
| progress_bar.close() | ||
|
|
||
|
|
||
| def iterate_file( |
Collaborator
There was a problem hiding this comment.
Have iterate_file take in epoch
See the following PR for reference
| yield _parse_jsonl_line(line) | ||
|
|
||
|
|
||
| async def train_sft_from_file( |
Collaborator
There was a problem hiding this comment.
Modify this so user can have the training continue running after closing their laptop.
- Iterate_file(file, epoch)
- Write to local disk
- Upload to wandb artifact
- Calculate lr
- Call train_sft(url, lr)
- Monitor training status
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.