Function build_train_sequences

general-knowledge/src/utils.py:193–237

Turn a raw completion + article context into the list of sequences that the inner-loop fine-tuning consumes. - `---` splits into separate training examples - if `split_newlines`, each line inside a segment becomes its own example - the original article context is always app

(
    completion_raw: str,
    context: str,
    title: str,
    *,
    split_newlines: bool = False,
    add_context: bool = True,
)

MAX_TRAIN_SEQS_PER_COMPLETION=30

def build_train_sequences(
    completion_raw: str,
    context: str,
    title: str,
    *,
    split_newlines: bool = False,
    add_context: bool = True,
) -> List[str]:
    """
    Turn a raw completion + article context into the list of sequences
    that the inner-loop fine-tuning consumes.

    - `---` splits into separate training examples  
    - if `split_newlines`, each line inside a segment becomes its own example  
    - the original article context is always appended as the last example
    - each example is prefixed with the title
    - if the second sequence begins with "1.", remove the first one
    """
    # For chain-of-thought, keep only content after the first "\nImplications:" marker
    m = re.search(r"\nImplications:\s*", completion_raw)
    if m:
        completion_raw = completion_raw[m.end():].lstrip()

    segs = _split_segments(completion_raw) or [completion_raw.strip()]
    if split_newlines:
        if re.search(r'Question\s+\d+:', completion_raw) and re.search(r'Answer\s*:', completion_raw): 
            # deal with self-QA responses
            # split wherever a new "Question N:" begins
            segs = re.split(r'\n(?=Question\s+\d+:)', completion_raw.strip())
            # ensure the very first segment has an explicit "Question 1:" prefix
            if not segs[0].lstrip().startswith("Question"):
                segs[0] = "Question 1: " + segs[0].strip()
        else: 
            # deal with responses where first line is along the lines of "Sure, let's give a list of implications:"
            segs = [ln.strip() for seg in segs for ln in seg.splitlines() if ln.strip()]
            if len(segs) > 1 and segs[1].startswith("1."):
                segs = segs[1:]

    if len(segs) > MAX_TRAIN_SEQS_PER_COMPLETION:
        segs = segs[:MAX_TRAIN_SEQS_PER_COMPLETION]
    segs = [s for s in segs if s.strip()]
    seqs = [TRAINING_SEQUENCE_TEMPLATE.format(title=title, completion_text=s) for s in segs]
    if add_context:
        seqs.append(TRAINING_SEQUENCE_TEMPLATE.format(title=title, completion_text=context.strip()))
    return seqs

# -------------------  PROXY GRADING  ---------------------------------- #
# This is a proxy used to evaluate the quality of synthetic data generated by the model.