Niccolò Zanichelli

I'm also encountering this issue. Line 383, which is now the same line as what (in @llwx593's screenshot over in #113) was line 380. StopIteration exception. My mmcif_dir looks like...

Thank you so much for the quick response Gustaf. You're right -- 57 chains appear to be missing. I'm unsure what caused it. The missing chains are `1erk 3d6s 3rvw...

Thank you again Gustaf for the quick response. I moved all the subdirectories to a separate "missing_mmcif" directory as to make sure no data was lost. I tried rerunning it...

Hi, thanks for the note. It looks like I had a typo. I'm training right now with a fixed config (`lora_rank==64`) and will close the issue if this is resolved...

I have retrained the model, this time with `parser.add_argument("--lora_rank", type=int, default=64)`. Now I have the adapted as well, as expected: `ls ckpt/qwen-7b-rm-highest-number/ README.md adapter_model.safetensors merges.txt model.safetensors.index.json tokenizer_config.json adapter_config.json added_tokens.json model-00001-of-00002.safetensors...

Training with the same config but with ZeRo stage 2 allows reward model checkpoint loading via `get_llm_for_sequence_regression`, but it warns > Some weights of LLMForSequenceRegression were not initialized from the...

I am now retrying with Llama 2 7B and zero stage 2. I'll see if I'm able to load the RM (including the value head) without issue and report back...

OK, I was able to train Llama 2 without issue, and here too I am able to load it as part of `get_llm_for_sequence_regression`: `model = get_llm_for_sequence_regression("ckpt/llama-2-7b-rm-zero-stage-2", model_type='reward', init_value_head=True)` However, I...

Hi, thanks for the note. Still, I'm not sure I understand. So, based on what you said, what's the proper way to load an _existing_ checkpoint of an _already finetuned_...

Fantastic, thank you. I'll test with an RM I'm training and close the issue when I'm able to confirm things are working as expected. Thanks again!