Datasets:

Modalities:
Text
Formats:
json
Size:
< 1K
Libraries:
Datasets
Dask
License:

DBRX-Instruct evaluation failed, likely due to model size (132B params)

#121
by abhi-db - opened

Hi! We recently merged the official DbrxModel implementation to transformers>=4.40.0 and the HF Hub models for databricks/dbrx-base, databricks/dbrx-instruct have been updated accordingly.

I tried to submit an eval request for databricks/dbrx-instruct but it failed for some reason. I suspect the eval hardware needs to be bigger to accomodate the large model size (132B params). It should definitely work on a 8x80GB system with bfloat16 precision, that's how I've been running evals locally.

If you could help eval this model manually we would very much appreciate it! When I ran eval locally I got an openllm average of around 74.5%, around the same as Command-R+.

https://huggingface.co/datasets/open-llm-leaderboard/requests/blob/main/databricks/dbrx-instruct_eval_request_False_bfloat16_Original.json

Hi @abhi-db !

According to the logs, the model failed to upload due to upload issues, but we have resubmitted your model and the evaluation should be successful now

I'll close this issue, please write here if you encounter any other problems with this model

alozowski changed discussion status to closed

Hi @alozowski , I think the resubmission for dbrx-instruct failed: https://huggingface.co/datasets/open-llm-leaderboard/requests/blob/main/databricks/dbrx-instruct_eval_request_False_bfloat16_Original.json

{
  "model": "databricks/dbrx-instruct",
  "base_model": "",
  "revision": "main",
  "private": false,
  "precision": "bfloat16",
  "params": 131.597,
  "architectures": "DbrxForCausalLM",
  "weight_type": "Original",
  "status": "FAILED",
  "submitted_time": "2024-04-19T07:36:33Z",
  "model_type": "\ud83d\udcac : chat models (RLHF, DPO, IFT, ...)",
  "job_id": "4054064",
  "job_start_time": "2024-04-29T11:57:23.745963"
}

Could you help take a look at the logs?

alozowski changed discussion status to open

Hi @abhi-db !

Unfortunately, according to the logs, the model failed due to a job preemption – sometimes our research cluster can be full. I've resubmitted your model, please, check out the request file. If the model fails again, please, ping me here and I will run manual evaluation 👍

alozowski changed discussion status to closed

Sign up or log in to comment