Got stuck when evaluating MMLU

[zhangliang-04]

Thanks for your open sourcing! i'm trying to evaluate Llama-7b-hf on mmlu-fr, a warning of Token indices sequence length is longer than the specified maximum sequence length for this model (5023 > 4096). Running this sequence through the model will result in indexing errors occurs and it seems the process is stuck. Here is the callstack after keyboard interrupt:

Token indices sequence length is longer than the specified maximum sequence length for this model (5023 > 4096). Running this sequence through the model will result in indexing errors
^CTraceback (most recent call last):
  File "/data2/zl/code/mlmm-evaluation/main.py", line 135, in <module>
    main()
  File "/data2/zl/code/mlmm-evaluation/main.py", line 108, in main
    results = evaluator.open_llm_evaluate(
  File "/data2/zl/code/mlmm-evaluation/lm_eval/utils.py", line 205, in _wrapper
    return fn(*args, **kwargs)
  File "/data2/zl/code/mlmm-evaluation/lm_eval/evaluator.py", line 79, in open_llm_evaluate
    results = evaluate(
  File "/data2/zl/code/mlmm-evaluation/lm_eval/utils.py", line 205, in _wrapper
    return fn(*args, **kwargs)
  File "/data2/zl/code/mlmm-evaluation/lm_eval/evaluator.py", line 262, in evaluate
    resps = getattr(lm, reqtype)([req.args for req in reqs])
  File "/data2/zl/code/mlmm-evaluation/lm_eval/base.py", line 181, in loglikelihood
    context_enc = self.tok_encode(context)
  File "/data2/zl/code/mlmm-evaluation/lm_eval/models/huggingface.py", line 361, in tok_encode
    return self.tokenizer.encode(string, add_special_tokens=self.add_special_tokens)
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_base.py", line 2569, in encode
    encoded_inputs = self.encode_plus(
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_base.py", line 2977, in encode_plus
    return self._encode_plus(
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_fast.py", line 576, in _encode_plus
    batched_output = self._batch_encode_plus(
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_fast.py", line 504, in _batch_encode_plus
    encodings = self._tokenizer.encode_batch(
KeyboardInterrupt

it seems the process is stuck in the batched tokenizing, how to deal with this?

[omarnj-lab]

Is your problem solved ? please let me know since I am dealing with same issue

[usernameisintentionallyhidden]

You can limit the #token feed to the model to match the max token length. This is rather a problem with lm_eval_harness than additional tasks.

[dame-cell]

You can limit the #token feed to the model to match the max token length. This is rather a problem with lm_eval_harness than additional tasks.

How do we do this exactly?

[dame-cell]

Thanks for your open sourcing! i'm trying to evaluate Llama-7b-hf on mmlu-fr, a warning of Token indices sequence length is longer than the specified maximum sequence length for this model (5023 > 4096). Running this sequence through the model will result in indexing errors occurs and it seems the process is stuck. Here is the callstack after keyboard interrupt:

Token indices sequence length is longer than the specified maximum sequence length for this model (5023 > 4096). Running this sequence through the model will result in indexing errors
^CTraceback (most recent call last):
  File "/data2/zl/code/mlmm-evaluation/main.py", line 135, in <module>
    main()
  File "/data2/zl/code/mlmm-evaluation/main.py", line 108, in main
    results = evaluator.open_llm_evaluate(
  File "/data2/zl/code/mlmm-evaluation/lm_eval/utils.py", line 205, in _wrapper
    return fn(*args, **kwargs)
  File "/data2/zl/code/mlmm-evaluation/lm_eval/evaluator.py", line 79, in open_llm_evaluate
    results = evaluate(
  File "/data2/zl/code/mlmm-evaluation/lm_eval/utils.py", line 205, in _wrapper
    return fn(*args, **kwargs)
  File "/data2/zl/code/mlmm-evaluation/lm_eval/evaluator.py", line 262, in evaluate
    resps = getattr(lm, reqtype)([req.args for req in reqs])
  File "/data2/zl/code/mlmm-evaluation/lm_eval/base.py", line 181, in loglikelihood
    context_enc = self.tok_encode(context)
  File "/data2/zl/code/mlmm-evaluation/lm_eval/models/huggingface.py", line 361, in tok_encode
    return self.tokenizer.encode(string, add_special_tokens=self.add_special_tokens)
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_base.py", line 2569, in encode
    encoded_inputs = self.encode_plus(
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_base.py", line 2977, in encode_plus
    return self._encode_plus(
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_fast.py", line 576, in _encode_plus
    batched_output = self._batch_encode_plus(
  File "/opt/conda/envs/lm_eval/lib/python3.10/site-packages/transformers/tokenization_utils_fast.py", line 504, in _batch_encode_plus
    encodings = self._tokenizer.encode_batch(
KeyboardInterrupt

it seems the process is stuck in the batched tokenizing, how to deal with this?

Did you fix this ??