Post

Reading the TÜLU 3 paper from @ai2.bsky.social. It's refreshing to see a research lab treating AI as a real science with full reports, data, code, logs, evals. Paper: allenai.org/papers/tulu-... Demo: playground.allenai.org Code: github.com/allenai/open... Eval: github.com/allenai/olmes Notes

Core skills - they identify some key areas open post-training models struggle with and target those for improving, picking datasets in those areas for evaluation. Notably, they avoid looking at scores from the unseen set during development to assess if their choices might have led to overfitting.

Data sources - they curate from public datasets based on three criteria - diversity (for generalization), targeting their identified core skills and clear licensing/provenance for proper usage. Datasets they considered: docs.google.com/spreadsheets...

Data sources - they combat mode collapse using persona-driven synthetic data generation, where unique personas (e.g., "ML researcher focused on neural networks") guide LLMs to create diverse training data for specific skills like precise instruction following, math, and coding.

Data decontamination - n-gram matching works better than embedding methods in terms of precision - 8-gram matching only on user-turns (since completions are LLM generated) - remove training sets that have more than 2% overlap with any evaluation set

SFT - for the data, train models on skill specific data mixtures, keep the mixtures that lead to the best performance on those skills. Later, combine the mixtures and do decontamination/downsampling of larger datasets.

SFT - a couple of data ablations: - Adding safety data didn't negatively affect other skills. - Measured the removal of math data on GSM8K and MATH, confirming that the scores drop. - Measured the the effectiveness of WildChat (diverse chat data) and Persona (skill based data)

SFT - using a better pretrained model (Qwen 2.5) results in better scores for GSM8K and MATH. Kinda surprising since the final model is still Llama 3.1, wonder if they evaluated Qwen 2.5 on other datasets as well and if the performance was not good enough to be the base model?

SFT - performance varies greatly depending on the random seed. They tried model soups (averaging weights of multiple models using github.com/arcee-ai/mer...), but decided to choose a single best model instead.

DPO - preference data is made of a on-policy model pool (Tulu 3 7B/70B) and off-policy model pool (other models). GPT-4o is used to rate generations from 4 random models from the model pool from 1-5. The highest is taken as the chosen response and the rejected response is sampled from the rest.

It's interesting how well LLM-as-a-judge worked for them here, it's completely viable now to evaluate your data using LLMs (even for seemingly fuzzy tasks like preference). related: aligned.substack.com/p/crisp-and-...

DPO ablations - the number of unique prompts in the preference dataset matter. More unique prompts result in better downstream performance, while more generations for the same prompts does not. Using different prompts from SFT data also improve performance.

DPO ablations - on-policy data (generations from the base SFT model) improve model performance. Although, it's interesting to note that they don't seem to train the model for multiple rounds like Llama 3.1 (maybe cost concerns?)

DPO ablations - there is much focus on ensuring model follows instruction, mostly leveraging synthetic data methods to create {chosen, rejected} pairs. They also measure that you can take an existing dataset, regenerate it using their synthetic pipeline and it improves performance. Really cool.

DPO ablations - doing PPO is hard, primarily because of the reward model, which is hard to evaluate. They conclude that **for the budget** DPO works well enough for preference tuning.

DPO ablations - reducing memory requirements by caching log probs across the dataset instead of computing on the fly with a reference model and separating forward passes for chosen and rejected sequences.

PPO - replaces the standard reward model with a verification function when the answer can be verified correct (e.g., in mathematics). The model is trained using PPO, only getting a reward if the answer is correct. This works really well to improve domains which have verifiable answers.

OLMES - they release the Open Language Model Evaluation System which is an evaluation suite for datasets similar to lm-evaluation-harness, but with more support for datasets, configurations, detailed output data for analysis. github.com/allenai/olmes

They also introduce more test datasets which use LLM-as-a-Judge - really seeming to be a common theme here. HREF uses human reference as a guide for LLM-as-a-Judge, allowing for a test set that can be used for open ended tasks such as summarization and rewriting.

And I really like this section: what didn't work. Both online DPO (training a reward model is hard) and rejection sampling (sampling n responses, choosing the best and using the rest as rejected) didn't improve performance.