shinnosukeono/alpaca_farm_preprocessing_1771395088

Alpaca Farm SFT Dataset Preprocessing

This dataset is preprocessed from tatsu-lab/alpaca_farm (alpaca_instructions subset) for use in RLHF pipelines.

Original Dataset Columns

• instruction: The task instruction
• input: Optional additional input/context
• output: The reference response

Added Preprocessing Columns

• query: Formatted prompt using Alpaca format (handles empty input)
• query_token: Tokenized query with padding (length=512)
• reference_response: Reference response with EOS token
• reference_response_token: Tokenized response (max_length=128)
• reference_response_token_len: Original length of tokenized response
• query_reference_response: Concatenation of query.strip() and reference_response
• query_reference_response_token: Tokenized concatenation (max_length=640)
• query_reference_response_token_response_label: Same as above but with query tokens masked for loss
• query_reference_response_token_len: Original length of concatenation

Prompt Format

### Instruction:
{instruction}

### Input:
{input}

### Response:

For examples with empty input:

### Instruction:
{instruction}

### Response:

Splits

• train: Training data
• validation: Validation data (10% of original)
• test: Test data (10% of original)

Args

{'alpaca_params': TaskQueryHParams(length=512,
                                   format_str='### Instruction:\n'
                                              '{instruction}\n'
                                              '\n'
                                              '### Input:\n'
                                              '{input}\n'
                                              '\n'
                                              '### Response:',
                                   truncate_field='input',
                                   truncate_text='\n',
                                   padding='pad_token',
                                   pad_token=[50277],
                                   pad_side='left',
                                   max_sft_response_length=128,
                                   max_sft_query_response_length=640),
 'base_model': 'EleutherAI/pythia-1b-deduped',
 'check_length_correctness': False,
 'debug': False,
 'hf_entity': 'shinnosukeono',
 'local_path': '/work/gj25/k36132/datasets/alpaca_preprocessed_pythia',
 'push_to_hub': True,
 'query_length': None,
 'response_length': None,
 'save_local': True}