chat_sl

Supervised Learning

SFT on NoRobots

python -m tinker_cookbook.recipes.chat_sl.train
    model_name=Qwen/Qwen3-8B-Base \
    dataset=no_robots \
    learning_rate=5e-4 \
    batch_size=64 \
    lora_rank=64 \
    eval_every=20 \
    save_every=20 \
    wandb_project=cookbook_sl

After 140 steps of training, test/nll decreases to 1.788.

SFT on Tulu3 dataset

python -m tinker_cookbook.recipes.chat_sl.train
    model_name=Qwen/Qwen3-8B-Base \
    dataset=tulu3 \
    learning_rate=5e-4 \
    batch_size=128 \
    lora_rank=64 \
    eval_every=500 \
    save_every=500 \
    wandb_project=cookbook_sl

After 1740 steps of training, test/nll decreases to 0.50. Performance can be further improved by training longer with a higher lora_rank and lower batch_size.

Adding your own dataset

The base classes in tinker_cookbook/supervised/data.py support loading new data in the following way:

• SupervisedDatasetFromHFDataset loads dataset on huggingface hub with a postprocessing function
• StreamingSupervisedDatasetFromHFDataset works similarly, but supports streaming
• FromConversationFileBuilder supports data loading from a JSONL file