SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2104.08661 · EMNLP · 2021

Explaining Answers with Entailment Trees

Oyvind Tafjord, Peter Clark, Peter Jansen, Bhavana Dalvi, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura

arXiv · PDF · Open in the Atlas

Code that ran

16 of 17 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 14 have no contract verdict; 2 failures of our instrument; 1 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2104.08661.json · badge: /api/ran/2104.08661.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
allenai/entailment_bank community (classifier label)
classifier basis not recorded on the link
16 ran · 1 no recorded run
FunctionStatusWhere it lives
count_phrase_matches Ran
ran on a synthesized input
allenai/entailment_bank/utils/entail_trees_utils.py
code served (permissive licence) · get_code("44b66cb0347665db")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def count_phrase_matches(phrase, reference):
    regex = '(\\b|(?!\\w))' + re.escape(phrase) + '((?<!\\w)|\\b)'
    return len(re.findall(regex, reference))

# Setup valid arguments
phrase = "hello"
reference = "hello world, hello there! hello-friend and hello_user"

ARGS = {
    "phrase": phrase,
    "reference": reference
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
decompose_slots Ran
ran on a synthesized input
allenai/entailment_bank/utils/nlp_agent.py
code served (permissive licence) · get_code("a09d8995568f8d93")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

DEFAULT_SLOT_FORMAT = {'slot': '$SLOT$', 'assign': ' = ', 'separator': ' ; ', 'missing_value': 'N/A'}

def decompose_slots(string, fmt=None):
    fmt = fmt or DEFAULT_SLOT_FORMAT
    string = string.strip()
    no_slot = 'PREFIX'
    slot_re = re.compile('(?i)' + re.escape(fmt['slot']).replace('SLOT', '(\\w*?)'))
    assign_re = re.escape(fmt['assign']).replace('\\ ', '\\s*')
    separator_re = re.escape(fmt['separator']).replace('\\ ', '\\s*')
    strip_re = re.compile(f'^({assign_re})?(.*?)({separator_re})?$')
    slot_pos = []
    for m in slot_re.finditer(string):
        slot_pos.append((m.span(), m.group(1)))
    if len(slot_pos) == 0:
        return {no_slot: string}
    if slot_pos[0][0][0] > 0:
        slot_pos = [((0, -1), no_slot)] + slot_pos
    res = {}
    for idx, (pos, slot_name) in enumerate(slot_pos):
        if idx == len(slot_pos) - 1:
            value = string[pos[1] + 1:]
        else:
            value = string[pos[1] + 1:slot_pos[idx + 1][0][0] - 1]
        m = strip_re.match(value)
        if m is not None:
            value = m.group(2)
        value = value.strip()
        if slot_name in res:
            value = res[slot_name] + ' ~AND~ ' + value
        res[slot_name] = value
    return res

ARGS = {
    'string': 'prefix text $SLOT$ = value1 ; $SLOT$ = value2 ; ',
    'fmt': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_jsonl Ran
ran on a synthesized input
allenai/entailment_bank/utils/angle_utils.py
code served (permissive licence) · get_code("c30d329a797e15bd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary JSONL file with valid test data
temp_dir = tempfile.mkdtemp()
test_file = os.path.join(temp_dir, "test.jsonl")

# Write some valid JSONL data
test_data = [
    {"id": 1, "text": "hello world"},
    {"id": 2, "text": "foo bar"},
    {"id": 3, "name": "test", "value": 42}
]

with open(test_file, 'w') as f:
    for item in test_data:
        f.write(json.dumps(item) + '\n')

ARGS = {
    "file_name": test_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_slot_data_by_id Ran
ran on a synthesized input
allenai/entailment_bank/utils/eval_utils.py
code served (permissive licence) · get_code("51014481762b0d1b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary file with JSON lines containing slot data
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.jsonl')
try:
    # Write some sample slot data
    slot_data = [
        {'id': 'slot_1', 'name': 'slot_a', 'value': 100},
        {'id': 'slot_2', 'name': 'slot_b', 'value': 200},
        {'id': 'slot_3', 'name': 'slot_c', 'value': 300},
    ]
    for item in slot_data:
        temp_file.write(json.dumps(item) + '\n')
    temp_file.close()
    
    slot_file_path = temp_file.name
    
    ARGS = {
        'slot_file': slot_file_path
    }
finally:
    # Note: temp file will persist for the function to read
    pass
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_predslot_record Ran
ran on a synthesized input
allenai/entailment_bank/eval/run_scorer_task3.py
code served (permissive licence) · get_code("213992abc08e4d79")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# Construct the pred dict with worldtree_provenance
pred = {
    'worldtree_provenance': {
        'fact1': {'original_text': 'The Earth orbits the Sun'},
        'fact2': {'original_text': 'Plants need sunlight'},
        'fact3': {'original_text': 'Water is essential for life'}
    },
    'prediction': 'some prediction text',
    'score': 0.95
}

# Construct the hypothesis string
hypothesis = 'Plants need sunlight to grow'

ARGS = {
    'pred': pred,
    'hypothesis': hypothesis
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
new_dict_update Ran
ran on a synthesized input
allenai/entailment_bank/utils/nlp_agent.py
code served (permissive licence) · get_code("baf18760bfe228c7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Analyze the function:
# - new_dict_update takes two parameters: old_dict and update_dict
# - Both should be dictionaries (or None for update_dict)
# - The function copies old_dict, updates it with update_dict if not None, and returns it

# Create test dictionaries
old_dict = {
    'learning_rate': 0.001,
    'batch_size': 32,
    'epochs': 10,
    'model_type': 'transformer'
}

update_dict = {
    'learning_rate': 0.0005,
    'epochs': 20,
    'new_param': 'added_value'
}

ARGS = {
    'old_dict': old_dict,
    'update_dict': update_dict
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_lisp Ran
ran on a synthesized input
allenai/entailment_bank/utils/proof_utils.py
code served (permissive licence) · get_code("970c22892bfbf76b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 2, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def parse_lisp(input_string, token_map=None):
    token_regex = re.compile('(->|[a-zA-Z0-9%@^]+|\\S)')
    tokens = token_regex.findall(input_string)
    if token_map is not None:
        tokens = [token_map.get(x, x) for x in tokens]
    return parse_lisp_tokens(tokens)

def parse_lisp_tokens(tokens):
    res = []
    i = 0
    while i < len(tokens):
        symbol = tokens[i]
        if symbol == '(':
            sub_list = []
            depth = 1
            while depth > 0:
                i += 1
                if i >= len(tokens):
                    raise ValueError('Invalid input: Unmatched open parenthesis.')
                symbol = tokens[i]
                if symbol == '(':
                    depth += 1
                elif symbol == ')':
                    depth -= 1
                if depth > 0:
                    sub_list.append(symbol)
            res.append(parse_lisp_tokens(sub_list))
        elif symbol == ')':
            raise ValueError('Invalid input: Unmatched close parenthesis.')
        else:
            res.append(symbol)
        i += 1
    return res

# Test case 1: Simple string with nested parentheses
input_string_1 = "(+ 1 2)"
token_map_1 = None

# Test case 2: String with token mapping
input_string_2 = "(func a b)"
token_map_2 = {"func": "->", "a": "x", "b": "y"}

# Test case 3: More complex nested structure
input_string_3 = "(+ (* 2 3) (- 5 1))"
token_map_3 = None

ARGS = {
    "input_string": "(+ 1 2)",
    "token_map": None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_lisp_tokens Ran
ran on a synthesized input
allenai/entailment_bank/utils/proof_utils.py
code served (permissive licence) · get_code("af3e6075f5840574")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

# Setup code for parse_lisp_tokens

ARGS = {
    'tokens': ['(', 'define', 'x', '5', ')', 'y', '(', '+', '1', '2', ')']
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
recursive_polish_notation Ran
ran on a synthesized input
allenai/entailment_bank/utils/proof_utils.py
code served (permissive licence) · get_code("b015b685dfa43094")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

# Setup code for recursive_polish_notation function

SYMBOL_DICT = {'->': '#', 'and': '&'}

def dedupe(data):
    res = []
    for d in data:
        if d not in res:
            res.append(d)
    return res

def recursive_polish_notation(proof_list):
    res = []
    if isinstance(proof_list, str):
        res.append(proof_list)
        return res
    if len(proof_list) == 3 and proof_list[1] == '->':
        res.append(SYMBOL_DICT['->'])
        res += recursive_polish_notation(proof_list[2])
        res += recursive_polish_notation(proof_list[0])
        return res
    proof_list = dedupe(proof_list)
    if len(proof_list) == 1:
        res += recursive_polish_notation(proof_list[0])
    elif len(proof_list) == 2:
        res.append(SYMBOL_DICT['and'])
        res += recursive_polish_notation(proof_list[0])
        res += recursive_polish_notation(proof_list[1])
    elif len(proof_list) > 2:
        res.append(SYMBOL_DICT['and'])
        res += recursive_polish_notation(proof_list[0])
        res += recursive_polish_notation(proof_list[1:])
    else:
        raise ValueError('Empty list found')
    return res

# Test case 1: Simple string input
test_input_1 = "A"

# Test case 2: Implication structure [A, '->', B]
test_input_2 = ["A", "->", "B"]

# Test case 3: List with multiple elements (and structure)
test_input_3 = ["A", "B", "C"]

# Test case 4: Nested structure with implication
test_input_4 = [["A", "->", "B"], "->", "C"]

# Test case 5: List with duplicates that will be deduplicated
test_input_5 = ["A", "B", "A"]

ARGS = {
    'proof_list': test_input_2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
replace_punctuation Ran
ran; our contract draft was wrong, not the code
allenai/entailment_bank/utils/eval_utils.py
code served (permissive licence) · get_code("d2166ef001a9e416")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

ARGS = {
    "str": 'Hello "world" and \'test\' string'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
score_string_similarity Ran
ran on a synthesized input
allenai/entailment_bank/utils/eval_utils.py
code served (permissive licence) · get_code("00139853f0c3a89e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

T5_GOOD_CHARS = [32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 93, 95, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 124, 163, 171, 173, 174, 176, 187, 201, 206, 220, 223, 224, 225, 226, 228, 231, 232, 233, 234, 238, 243, 244, 246, 249, 251, 252, 259, 351, 355, 537, 539, 1072, 1074, 1076, 1077, 1080, 1082, 1083, 1084, 1085, 1086, 1088, 1089, 1090, 1091, 8211, 8212, 8216, 8217, 8220, 8221, 8222, 8226, 8242, 8364, 9601]

T5_BAD_REGEX = re.compile('[^' + re.escape('.'.join([chr(x) for x in T5_GOOD_CHARS])) + ']')

def fix_t5_unk_characters(str):
    return re.sub(' {2,}', ' ', re.sub(T5_BAD_REGEX, ' ', str))

def score_string_similarity(str1, str2):
    if str1 == str2:
        return 3.0
    str1 = fix_t5_unk_characters(str1)
    str2 = fix_t5_unk_characters(str2)
    if str1 == str2:
        return 2.0
    str1 = str1.lower()
    str2 = str2.lower()
    if str1 == str2:
        return 1.5
    if ' ' in str1 or ' ' in str2:
        str1_split = str1.split(' ')
        str2_split = str2.split(' ')
        overlap = list(set(str1_split) & set(str2_split))
        return len(overlap) / max(len(str1_split), len(str2_split))
    else:
        return 0.0

ARGS = {
    "str1": "Hello world test",
    "str2": "Hello world example"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sentence_index Ran
ran on a synthesized input
allenai/entailment_bank/utils/entail_trees_utils.py
code served (permissive licence) · get_code("8f025f3f4d630f9a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

def sentence_index(sentence_id):
    return int(sentence_id.replace('sent', ''))

ARGS = {
    'sentence_id': 'sent42'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sort_sentence_ids_match Ran
ran on a synthesized input
allenai/entailment_bank/utils/entail_trees_utils.py
code served (permissive licence) · get_code("bc66cae80c3a0886")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

class MockMatch:
    def __init__(self, text):
        self.text = text
    
    def group(self, index):
        return self.text

ARGS = {
    'match': MockMatch('sent3 & sent1 & sent2')
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
split_info_sentences Ran
ran on a synthesized input
allenai/entailment_bank/eval/run_scorer.py
code served (permissive licence) · get_code("110b38cab88d9900")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def split_info_sentences(context):
    words_list = context.split(' ')
    sentence_ids = re.findall('[\\w\\.]+[0-9]+:', context)
    sentence_dict = dict()
    prev_sid = ''
    prev_sentence_parts = []
    for word in words_list:
        if word in sentence_ids:
            if prev_sid:
                sentence_dict[prev_sid] = ' '.join(prev_sentence_parts)
                prev_sentence_parts = []
            prev_sid = word
        else:
            prev_sentence_parts.append(word)
    if prev_sid:
        sentence_dict[prev_sid] = ' '.join(prev_sentence_parts)
    return sentence_dict


# The function takes a context string and parses it
# It looks for sentence IDs (pattern: word characters/dots followed by digits and colon)
# and groups subsequent words into those sentence IDs

context = "sent1: this is the first sentence sent2: this is the second sentence"

ARGS = {
    "context": context
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
split_mcoptions Ran
ran on a synthesized input
allenai/entailment_bank/utils/nlp_agent.py
code served (permissive licence) · get_code("6cfa0a08a9e5f18a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def split_mcoptions(mcoptions):
    first_option = ord(mcoptions.strip()[1])
    labels = ''.join([chr(x) for x in range(first_option, first_option + 10)])
    choices = re.split('\\s*\\([' + labels + ']\\)\\s*', mcoptions)[1:]
    return (choices, chr(first_option))

# Construct valid arguments for split_mcoptions
# The function expects a string with:
# - Some leading whitespace or character at position 0
# - A character at position 1 that will be the first option
# - Text containing patterns like "(a) choice1 (b) choice2" etc.

mcoptions_str = " a) First choice (b) Second choice (c) Third choice"

ARGS = {
    "mcoptions": mcoptions_str
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
trim_batch Ran
ran; our contract draft was wrong, not the code
allenai/entailment_bank/utils/angle_utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Sahandfer/CEM, is pointer only (licence: NONE) · get_code("eeb8dc60c6ef1fa4")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Sahandfer/CEM, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

encode_line Not yet run
unverified (harvested, no recorded run)
allenai/entailment_bank/utils/angle_utils.py
code served (permissive licence) · get_code("ab5d9464aeabf9fd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 17 samples lifted from this paper's linked repositories, 17 carry a function identity and 0 do not. Of the 17 distinct functions with an identity, 13 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 13 later) and 4 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
trim_batch 9 other papers
1 dated earlier, 8 later
2005.01107 (2020-05)
Simplifying Paragraph-level Question Generation via Transformer Language Models
2108.02866 (2021-08)
Dual Reader-Parser on Hybrid Textual and Tabular Evidence for Open Domain Question Answering
2204.06239 (2022-04)
Can Question Rewriting Help Conversational Question Answering?
and 6 more not named here
encode_line 4 other papers
4 later
2109.12595 (2021-09)
MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
2023.findings-acl.743 (2023 (year only))
KORC: Knowledge oriented Reading Comprehension Benchmark for Deep Text Understanding
and 1 more not named here
replace_punctuation 3 other papers
3 later
2202.12359 (2022-02)
UnifiedQA-v2: Stronger Generalization via Broader Cross-Format Training
2203.07540 (2022-03)
SCIENCEWORLD: Is your Agent Smarter than a 5 th Grader?
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
load_jsonl 2 other papers
2 later
2203.07540 (2022-03)
SCIENCEWORLD: Is your Agent Smarter than a 5 th Grader?
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
score_string_similarity 2 other papers
2 later
2203.07540 (2022-03)
SCIENCEWORLD: Is your Agent Smarter than a 5 th Grader?
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
count_phrase_matches 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
load_slot_data_by_id 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
parse_lisp 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
parse_lisp_tokens 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
recursive_polish_notation 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
sentence_index 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
sort_sentence_ids_match 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner
split_info_sentences 1 other paper
1 later
2205.09224 (2022-05)
Entailment Tree Explanations via Iterative Retrieval-Generation Reasoner

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Our goal, in the context of open-domain textual question-answering (QA), is to explain answers by showing the line of reasoning from what is known to the answer, rather than simply showing a fragment of textual evidence (a "rationale"). If this could be done, new opportunities for understanding and debugging the system's reasoning become possible. Our approach is to generate explanations in the form of entailment trees, namely a tree of multipremise entailment steps from facts that are known, through intermediate conclusions, to the hypothesis of interest (namely the question + answer). To train a model with this skill, we created ENTAILMENTBANK 1 , the first dataset to contain multistep entailment trees. Given a hypothesis (question + answer), we define three increasingly difficult explanation tasks: generate a valid entailment tree given (a) all relevant sentences (b) all relevant and some irrelevant sentences, or (c) a corpus. We show that a strong language model can partially solve these tasks, in particular when the relevant sentences are included in the input (e.g., 35% of trees for (a) are perfect), and with indications of generalization to other domains. This work is significant as it provides a new type of dataset (multistep entailments) and baselines, offering a new avenue for the community to generate richer, more systematic explanations.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2104.08661")
get_code_for_paper("2104.08661")
have("2104.08661")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2104.08661.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2104.08661.svg)](https://syntology.ai/paper/2104.08661)

Connect an agent — have() is free.