fix compatibility issues with nyuv2 experiments #47
Conversation
Co-authored-by: Copilot <[email protected]>
There was a problem hiding this comment.
Copilot reviewed 13 out of 27 changed files in this pull request and generated 3 suggestions.
Files not reviewed (14)
- config/llama_weighted_average.yaml: Language not supported
- examples/lm_finetune/llama_fullfinetune.sh: Language not supported
- fusion_bench/dataset/llama/collate.py: Evaluated as low risk
- fusion_bench/compat/modelpool/init.py: Evaluated as low risk
- fusion_bench/compat/taskpool/init.py: Evaluated as low risk
- config/method/lm_finetune/peftfinetune_sft.yaml: Evaluated as low risk
- config/modelpool/SeqenceClassificationModelPool/llama_preference700k.yaml: Evaluated as low risk
- fusion_bench/method/lm_finetune/bradley_terry_rm.py: Evaluated as low risk
- config/nyuv2_config.yaml: Evaluated as low risk
- config/modelpool/CausalLMPool/llama_ultrachat.yaml: Evaluated as low risk
- config/modelpool/SeqenceClassificationModelPool/single_reward_model.yaml: Evaluated as low risk
- config/dataset/llm_sft/ultrachat_200k.yaml: Evaluated as low risk
- fusion_bench/dataset/llama/preference_700k.py: Evaluated as low risk
- config/taskpool/reward_model_evaluation.yaml: Evaluated as low risk
Comments skipped due to low confidence (3)
fusion_bench/dataset/llama/ultrachat.py:37
- [nitpick] The comment on line 37 is unnecessary and should be removed for clarity.
# ? is it necessary to `.replace(tokenizer.bos_token, '')`?
fusion_bench/dataset/llama/stanford_shp.py:69
- The assertion error message uses an undefined variable 'positive'. It should use 'chosen' instead.
assert (tokenizer.eos_token_id not in tokenized_pos["input_ids"][:-1]), f"Prompt contains EOS token: {sample['positive']}"
fusion_bench/dataset/llama/stanford_shp.py:79
- The assertion error message uses an undefined variable 'rejected'. It should use 'rejected' instead.
assert (tokenizer.eos_token_id not in tokenized_neg["input_ids"][:-1]), f"Prompt contains EOS token: {sample['rejected']}"
| cache_path: Optional[str] = None, | ||
| ): | ||
| R""" | ||
| Load and tokenized Ultrachat 200k dataset for Bradley-Terry ranking model. |
There was a problem hiding this comment.
The docstring is misleading as it mentions 'winner' which is not relevant to the actual function. It should be updated to accurately describe the function's behavior.
| Load and tokenized Ultrachat 200k dataset for Bradley-Terry ranking model. | |
| Load and tokenize the Ultrachat 200k dataset for Bradley-Terry ranking model. |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
| sample["rejected_input_ids"].append(tokenizer.eos_token_id) | ||
| sample["rejected_attention_mask"].append(1) | ||
|
|
||
| dataset = dataset.map(tokenize, num_proc=num_proc) |
There was a problem hiding this comment.
The 'tokenize' function does not return the modified sample, which is required for 'dataset.map' to work correctly.
| dataset = dataset.map(tokenize, num_proc=num_proc) | |
| return sample |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
| @@ -0,0 +1,6 @@ | |||
| alpaca-cleaned: | |||
| _target_: fusion_bench.dataset.llama.alpaca.load_tokenized_alpaca_dataset | |||
| tokenizer: ??? | |||
There was a problem hiding this comment.
The tokenizer field is set to ???. It should be replaced with the appropriate tokenizer.
| tokenizer: ??? | |
| tokenizer: 'appropriate_tokenizer' |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
No description provided.