MIT-OL-AI-D/cv-nlp-lab2

CV & NLP Lab 2: language model systems

Teaching data for Lab 2 of the MIT CV & NLP MicroMasters course (6.4600.2x), "Language Model Systems". Learners prompt, fine-tune and quantize Qwen2.5-0.5B-Instruct on grade-school math problems and compare the systems on the same problems. The bundle holds the problems and the outputs of every system the lab compares, so a learner can read and score any system without rerunning it.

Files

File	Holds
lab2_data.json.gz	gzipped JSON: the problems and their splits, the solutions used for fine-tuning and calibration, every cached run, and the log-likelihoods of the gold solutions
reference_adapter/	a LoRA adapter for Qwen2.5-0.5B-Instruct, in the PEFT format, for a learner whose own training run fails

Key in lab2_data.json.gz	Holds
problems	1,319 GSM8K test problems by ID: question, written solution and final answer
full, test200, dev	the IDs of the 1,269 scored problems, the 200-problem panel and the 50 development problems
exemplars	the 4 worked examples the few-shot prompt shows
lora_train	330 GSM8K training problems with the model's own correct solution to each
calibration, calibration_pool	GSM8K training problems with their written solutions: the shared 128 and the pool of 500
runs	19 systems on the 1,269 problems: each output's raw text, prompt and output token counts, and whether it finished
runs_200, runs_dev	the prompting systems on the 200-problem panel and on the development problems
log_likelihood	the summed log-likelihood and token count of each gold solution, under the base model and two int4 versions
audit	the IDs of ten outputs the lab reads closely

Sources and licenses

• GSM8K (Cobbe et al. 2021, openai/grade-school-math): MIT license. Calculator annotations are stripped.
• Qwen2.5-0.5B-Instruct (Qwen team, Alibaba Cloud), revision 7ae557604adf67be50417f59c2c2f167def9a775: Apache 2.0. The model outputs and the adapter are derived from it.

How it was built

The problems are split with seed 6460. Every output was generated on an NVIDIA T4 during the course's pilot runs. The adapter has rank 16 and alpha 32 on every attention and MLP projection, and was trained for two epochs on the model's own correct solutions to the training problems, with seed 0.

Citations

@article{cobbe2021gsm8k,
  title   = {Training Verifiers to Solve Math Word Problems},
  author  = {Cobbe, Karl and Kosaraju, Vineet and Bavarian, Mohammad and Chen, Mark and Jun, Heewoo and
             Kaiser, Lukasz and Plappert, Matthias and Tworek, Jerry and Hilton, Jacob and Nakano, Reiichiro and
             Hesse, Christopher and Schulman, John},
  journal = {arXiv preprint arXiv:2110.14168},
  year    = {2021}
}
@article{qwen2.5,
  title   = {Qwen2.5 Technical Report},
  author  = {{Qwen Team}},
  journal = {arXiv preprint arXiv:2412.15115},
  year    = {2024}
}