first-sft-run.md

Run Your First SFT Job

The conversational SFT recipe is the current shortest end-to-end training path. It creates a training job, loads a chat dataset, submits forward/backward batches, applies optimizer steps, and cancels the job when complete.

Install the recipe dependency:

uv pip install 'tinker-cookbook @ git+https://github.com/thinking-machines-lab/tinker-cookbook.git@nightly'

Start with a short run:

python -m recipes.sft.conversational.train \
  config=/path/to/config.json \
  max_steps=2

The default job config, configs/qwen3_8b_full.json, requests four GPUs and runs full-parameter training, so check capacity first. GPU count, batch shape, sequence length and LoRA all live in that JSON rather than on the command line -- pass a different one with job_config=, for example job_config=configs/qwen3_8b_lora.json for the lighter LoRA path. See sizing and batching.

For a longer run, dataset changes, dense training, and MoE configuration, see the conversational SFT recipe.

There is no packaged before/after evaluation workflow yet. To confirm the run worked, check that train_mean_nll and test/nll fall over a longer run, then run the sampling command the recipe prints after it saves its checkpoint -- on the default memorize task the answer should be Snowflake AI Research.