Spaces:
Running
on
CPU Upgrade
What should a finetuned model's license be if the model is MIT but the datasets are Apache 2.0 and cc-by-4.0
I submitted an instruction tuned version of Microsoft's Phi 1.5 LLM, rasyosef/Phi-1_5-Instruct-v0.1 to the Open LLM Leaderboard last week, but it later disappeared from the Pending Evaluation Queue. Was it removed because of the datasets I used or some other reason?
The base model phi 1.5 has an MIT license but the datasets I used have Apache 2.0 and cc-by-4.0, which of these licenses would the finetuned rasyosef/Phi-1_5-Instruct-v0.1 model inherit?
Supervised Fine-Tuning
- teknium/OpenHermes-2.5 ~ MIT License
Direct Preference Optimization (DPO)
- Used a combination of the following preference datasets
- HuggingFaceH4/ultrafeedback_binarized ~ MIT License
- argilla/distilabel-intel-orca-dpo-pairs ~ Apache 2.0 License
- argilla/distilabel-math-preference-dpo ~ Apache 2.0 License
- jondurbin/py-dpo-v0.1 ~ cc-by-4.0 License
- Used a combination of the following preference datasets
I found the submission entry in the requests dataset and it shows the status as Failed.
{
"model": "rasyosef/Phi-1_5-Instruct-v0.1",
"base_model": "",
"revision": "69307c5e1ab1367aad818118253ddee7578f3c65",
"precision": "bfloat16",
"params": 1.415,
"architectures": "PhiForCausalLM",
"weight_type": "Original",
"status": "FAILED",
"submitted_time": "2024-07-25T01:54:51Z",
"model_type": "\ud83d\udcac : \ud83d\udcac chat models (RLHF, DPO, IFT, ...)",
"job_id": "7816415",
"job_start_time": "2024-07-29T08:27:00.038162",
"use_chat_template": true,
"sender": "rasyosef"
}
If possible I would like to know the reason. I apologize in advance in case the failure was caused by something on my end.
Hi!
Thanks for your issue!
For your model, it failed at loading - it's a bit hard to say if it was a hardware failure or model failure, so I'm passing it to pending again and we'll see how it relaunches.
For the license, hard to say! MIT and Apache are compatible together iirc, but not sure how that plays with a cc by.
Cool, Thanks!