Finbarr Timbers

pip support

1

Have you looked into making this available via pip? If not, I'm happy to look into contributing a PR.

Right now, the sidebar link points to https://pkomiske.github.io/Wasserstein/, which is down!

Add explicit dependency on flask_wtf 0.13.1.

9

## Description `flask_wtf` 0.14 introduces breaking changes. I believe this is because `flask_wtf` doesn't use `SecureForm` anymore. See: https://flask-wtf.readthedocs.io/en/stable/changelog.html#version-0-14 ## Steps to replicate 1. Update `flask_wtf` with `pip install --upgrade...

Struggling to replicate evaluation results

14

Hi folks, I'm trying to replicate your linear probe evaluation results. I can only get your pre-trained model to score 77% (with the last layer) or 80.8% (with the last...

Adds three scripts: 1. `scripts/regenerate_dataset_completions.py` which submits the batch job from a dataset. 2. `scripts/check_azure_batch_status.py` which prints the status of the batch job. 3. `scripts/process_azure_batch_results.py` which uploads the results to...

Now, `reward_fn` lives in LLMRayActor.

9

Runs: - Single GPU GRPO: [Beaker](https://beaker.allen.ai/orgs/ai2/workspaces/open-instruct-dev/work/01KAVC7VF1GPHC2H7TX5HKBNS6?taskId=01KAVC7VF4X3M0K0FEANCHCC5R&jobId=01KAVC7VJNTZYRJNHNT92HDF1E) - Multi-node GRPO: [Beaker](https://beaker.allen.ai/orgs/ai2/workspaces/open-instruct-dev/work/01KAVNYWVA1KRK9A1SZ8YSBKMH?taskId=01KAVNYWVEQBJZ9AXWXJ8JQKV3&jobId=01KAVNYWZ379GA9060BY1XRB9H)

Ran the single GPU GRPO script: [Beaker](https://beaker.allen.ai/orgs/ai2/workspaces/open-instruct-dev/work/01KBDGJDDBC4SY98DYXRFGEXD3?taskId=01KBDGJDDHZKTB056NJN53QDCK&jobId=01KBDGJDH2T4XWMBQ41NJRW8VP). I also verified manually that we get an error if someone tries to pass the old values in as a flag: ``` grpo_fast.py:...

Switches generation to use vllm's OpenAI API, rather than going through vLLM directly.

1

This will make it easier to switch out API providers in the future (e.g. SGLang) and enables us to use vLLM's native tool parsing (in a subsequent PR). Runs: -...

> [!NOTE] > Run `make style-check` in the Quality GitHub Actions workflow before `make quality-check`. > > - **CI / GitHub Actions**: > - Update `/.github/workflows/quality.yml` to run `make style-check`...

Refactors `PolicyTrainerRayProcess` so that the `train` method uses a dict of data

2

This lets us clean it up by using comprehensions and loops instead of making repeated calls. Also switches the variables in `train` to use Shazeer style [shape suffixes](https://medium.com/@NoamShazeer/shape-suffixes-good-coding-style-f836e72e24fd). Also switches...