vt_utils.py

# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# adapted from https://github.com/NVIDIA/RULER/blob/main/scripts/data/synthetic/variable_tracking.py

import itertools
import random
import string

import datasets
import numpy as np
from tqdm import tqdm

from lm_eval.tasks.ruler.common_utils import DEFAULT_SEQ_LENGTHS, get_tokenizer


TASKS = {
    "variable_tracking": {
        "tokens_to_generate": 30,
        "template": """Memorize and track the chain(s) of variable assignment hidden in the following text.\n\n{context}\nQuestion: Find all variables that are assigned the value {query} in the text above.""",
        "answer_prefix": """ Answer: According to the chain(s) of variable assignment in the text above, {num_v} variables are assgined the value {query}, they are: """,
    },
}

TEMPLATE = (
    TASKS["variable_tracking"]["template"] + TASKS["variable_tracking"]["answer_prefix"]
)


def generate_chains(
    num_chains: int, num_hops: int, is_icl: bool = False
) -> tuple[list[list[str]], list[list[str]]]:
    vars_all = []
    k = 5 if not is_icl else 3
    num_hops = num_hops if not is_icl else min(10, num_hops)
    vars_all = [
        "".join(random.choices(string.ascii_uppercase, k=k)).upper()
        for _ in range((num_hops + 1) * num_chains)
    ]
    while len(set(vars_all)) < num_chains * (num_hops + 1):
        vars_all.append("".join(random.choices(string.ascii_uppercase, k=k)).upper())

    vars_ret = []
    chains_ret = []
    for i in range(0, len(vars_all), num_hops + 1):
        this_vars = vars_all[i : i + num_hops + 1]
        vars_ret.append(this_vars)
        this_chain = [f"VAR {this_vars[0]} = {np.random.randint(10000, 99999)}"]
        for j in range(num_hops):
            this_chain.append(f"VAR {this_vars[j + 1]} = VAR {this_vars[j]} ")
        chains_ret.append(this_chain)
    return vars_ret, chains_ret


def generate_input_output(num_noises, num_chains, num_hops, is_icl=False):
    vars, chains = generate_chains(num_chains, num_hops, is_icl=is_icl)

    noise = "The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.\n"

    # Create a list of the repeated noise
    sentences = [noise] * num_noises
    if len(sentences) <= len(chains[0]):
        sentences = [
            n + "." if len(n.strip()) > 0 else n
            for n in [x for noise in sentences for x in noise.split(".")]
        ]
        try:
            assert len(sentences) > len(chains[0]), (
                "Noises too short, unable to generate data"
            )
        except:  # noqa: E722
            print("reduces chain length for not enough noises")
            chains = [chain[: len(sentences) - 1] for chain in chains]
    # sample random positions to insert variable assignment
    for chain_i in chains:
        # sample random positions (sorted) to insert variable assignment
        positions = list(sorted(random.sample(range(len(sentences)), len(chain_i))))
        for insert_pi, j in zip(positions, range(len(chain_i))):
            sentences.insert(insert_pi + j, chain_i[j])

    # Insert the passkey sentence at the random position
    context = " ".join(sentences)
    context = context.replace(". \n", ".\n")

    template = TEMPLATE
    if (
        is_icl
        and template
        != TASKS["variable_tracking"]["template"]
        + TASKS["variable_tracking"]["answer_prefix"]
    ):
        # remove model template
        cutoff = template.index(TASKS["variable_tracking"]["template"][:20])
        cutoff_ans = template.index(TASKS["variable_tracking"]["answer_prefix"][:10])
        template = (
            " ".join(template[cutoff:cutoff_ans].split()[:-1]) + template[cutoff_ans:]
        )

    value = chains[0][0].split("=")[-1].strip()
    input_text = template.format(context=context, query=value, num_v=num_hops + 1)

    return input_text, vars[0]


def randomize_icl(icl_example: str) -> str:
    icl_tgt_cut = icl_example.index(TASKS["variable_tracking"]["answer_prefix"][-10:])
    icl_tgt = icl_example[icl_tgt_cut + 10 :].strip().split()
    for item in icl_tgt:
        new_item = "".join(random.choices(string.ascii_uppercase, k=len(item))).upper()
        icl_example = icl_example.replace(item, new_item)
    return icl_example


def sys_vartrack_w_noise_random(
    tokenizer,
    num_samples: int,
    max_seq_length: int,
    incremental: int = 10,
    num_chains: int = 1,
    num_hops: int = 4,
    add_fewshot: bool = True,
    tokens_to_generate=30,
    icl_example: dict = None,
    remove_newline_tab=False,
):
    write_jsons = []
    tokens_to_generate = tokens_to_generate

    # Find the perfect num_noises
    num_noises = incremental

    total_tokens = 0  # Track the total tokens generated for this example
    example_tokens = 0
    if add_fewshot and (icl_example is not None):
        icl_example_out = " ".join(icl_example["outputs"])
        icl_example = icl_example["input"] + " " + icl_example_out + "\n\n"
        example_tokens = len(tokenizer(icl_example).input_ids)

    while total_tokens + tokens_to_generate + example_tokens < max_seq_length:
        input_text, answer = generate_input_output(
            num_noises, num_chains, num_hops, is_icl=add_fewshot & (icl_example is None)
        )
        # Calculate the number of tokens in the example
        total_tokens = len(tokenizer(input_text + f" {answer}").input_ids)
        print(
            f"Max length {max_seq_length} | Current length {total_tokens + tokens_to_generate + example_tokens} | Noises: {num_noises}"
        )
        if total_tokens + tokens_to_generate + example_tokens > max_seq_length:
            num_noises -= incremental
            break
        num_noises += incremental
    print("Num noises:", num_noises)

    # Generate samples
    for index in tqdm(range(num_samples)):
        used_noises = num_noises
        while True:
            try:
                input_text, answer = generate_input_output(
                    used_noises,
                    num_chains,
                    num_hops,
                    is_icl=add_fewshot & (icl_example is None),
                )
                length = (
                    len(tokenizer(input_text).input_ids)
                    + tokens_to_generate
                    + example_tokens
                )
                assert length <= max_seq_length, f"{length} exceeds max_seq_length."
                break
            except:  # noqa: E722
                if used_noises > incremental:
                    used_noises -= incremental

        if add_fewshot and (icl_example is not None):
            # insert icl_example between model template and input
            cutoff = input_text.index(TASKS["variable_tracking"]["template"][:20])
            input_text = (
                input_text[:cutoff]
                + randomize_icl(icl_example)
                + "\n\n"
                + input_text[cutoff:]
            )
        if remove_newline_tab:
            input_text = " ".join(
                input_text.replace("\n", " ").replace("\t", " ").strip().split()
            )

        gen_prefix_index = input_text.rfind(
            " Answer: According to the chain(s) of variable assignment"
        )
        gen_prefix = input_text[gen_prefix_index:].strip()
        # This condition is to check if we are generating the few-shot.
        if icl_example is not None:
            input_text = input_text[:gen_prefix_index]
        formatted_output = {
            "index": index,
            "input": input_text,
            "outputs": answer,
            "length": length,
            "max_length": max_seq_length,
            "gen_prefix": gen_prefix.strip(),
        }
        write_jsons.append(formatted_output)

    return write_jsons


def get_dataset(pretrained, seq=None, **kwargs) -> list[dict]:
    tokenizer = get_tokenizer(pretrained)
    icl_example = sys_vartrack_w_noise_random(
        tokenizer=tokenizer,
        num_samples=1,
        max_seq_length=500,
        incremental=5,
    )[0]
    write_jsons = sys_vartrack_w_noise_random(
        tokenizer=tokenizer,
        num_samples=500,
        max_seq_length=seq,
        icl_example=icl_example,
    )
    return write_jsons


def get_vt_dataset(**kwargs) -> dict[str, datasets.Dataset]:
    pretrained = kwargs.get("tokenizer", kwargs.get("pretrained", {}))
    df = (
        get_dataset(pretrained, seq=seq)
        for seq in kwargs.pop("max_seq_lengths", DEFAULT_SEQ_LENGTHS)
    )

    return {
        "test": datasets.Dataset.from_list(
            list(itertools.chain.from_iterable(df)), split=datasets.Split.TEST
        )
    }