lapras_dataloader.py

"""Processor and collator for the Lapras Reasoning Datasets, in the style of a VLM data pipeline."""

import torch


class TimeSeriesProcessor:
    """Tokenizes text and packs time series into tensors, as a VLM processor does for images."""

    def __init__(self, tokenizer):
        self.tokenizer = tokenizer
        self.tokenizer.padding_side = "right"

    def apply_chat_template(self, prompt, response=None):
        messages = [{"role": "user", "content": prompt}]
        text = self.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        return text if response is None else text + response + self.tokenizer.eos_token

    def __call__(self, text, timeseries):
        batch = self.tokenizer(text, padding=True, add_special_tokens=False, return_tensors="pt")
        n_ch = max(len(x) for x in timeseries)
        n_t = max(len(ch) for x in timeseries for ch in x)
        values = torch.zeros(len(timeseries), n_ch, n_t)
        mask = torch.zeros(len(timeseries), n_ch, n_t, dtype=torch.bool)
        for i, x in enumerate(timeseries):
            for c, ch in enumerate(x):
                values[i, c, : len(ch)] = torch.tensor(ch)
                mask[i, c, : len(ch)] = True
        batch["timeseries"] = values
        batch["timeseries_mask"] = mask
        return batch


class TimeSeriesCollator:
    """Builds a training batch in which only the response is supervised (labels are -100 elsewhere)."""

    def __init__(self, processor):
        self.processor = processor

    def __call__(self, examples):
        prompts = [self.processor.apply_chat_template(ex["input"]) for ex in examples]
        texts = [self.processor.apply_chat_template(ex["input"], ex["output"]) for ex in examples]
        batch = self.processor(texts, [ex["timeseries"] for ex in examples])
        labels = batch["input_ids"].clone()
        labels[batch["attention_mask"] == 0] = -100
        for i, prompt in enumerate(prompts):
            labels[i, : len(self.processor.tokenizer(prompt, add_special_tokens=False)["input_ids"])] = -100
        batch["labels"] = labels
        return batch