test_maze.py

import os
from model.transformer import GPTConfig, GPT
from model.transformer_rope import GPTRoPEConfig, GPTRoPE
from model.transformer_nextlat import TransformerNextLatConfig, TransformerNextLat
from model.mamba import MambaConfig, Mamba
from model.mamba2 import Mamba2Config, Mamba2
from model.gated_deltanet import GatedDeltaNetConfig, GatedDeltaNet
from model.gru import GRUConfig, GRU
import numpy as np
import networkx as nx
import argparse
import pickle
import re
import torch
import math
from torch.nn.utils.rnn import pad_sequence
from cli_utils import parse_count, format_count


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--ckpt_iter', type=int, default=10000)
    parser.add_argument('--model', type=str, default='transformer', choices=['transformer', 'transformer-rope', 'transformer-nextlat', 'mamba', 'mamba2', 'gated-deltanet', 'gru'],
                        help='Model architecture; selects out/<model>/ and how the checkpoint is built.')
    parser.add_argument('--config', type=str, default='12_12_576')
    parser.add_argument('--temperature', type=float, default=1)
    parser.add_argument('--device', type=str, default='cuda:0')
    parser.add_argument('--num_nodes', type=int, default=100)
    parser.add_argument('--num_of_paths', type=int, default=20)
    parser.add_argument('--batch_size', type=int, default=100)
    parser.add_argument('--num_iters', type=int, default=10)
    parser.add_argument('--multitasks', action=argparse.BooleanOptionalAction, default=True,
                        help='Use multitask data (default: True)')
    parser.add_argument('--num_train_dataset', type=parse_count, default='10M',
                        help='Number of multitask training entries (supports K/M/B, default: 50000)')
    parser.add_argument('--num_test_dataset', type=parse_count, default=10000,
                        help='Number of multitask test entries (supports K/M/B, default: 10000)')
    parser.add_argument('--tasks', type=str, default='H1',
                        help='Task specification (e.g., A1, A1B1, A3B2, A1D1F1). Default: A1')
    parser.add_argument('--CL', action=argparse.BooleanOptionalAction, default=False,
                        help='Task C turn-label mode (default: False)')
    parser.add_argument('--graph_file', type=str, default=None,
                        help='Optional GraphML path; if provided, load this graph instead of the default')
    parser.add_argument('--B_graph_file', type=str, default=None,
                        help='Optional GraphML path for Task B; if provided, use this graph for Task B validation instead of the default')
    parser.add_argument('--local', action='store_true', default=False,
                        help='Disable flash attention for local GPU compatibility (default: False)')
    parser.add_argument('--path_type', type=str, default='RWs', choices=['RWc', 'RWa', 'RWs'],
                        help='Path generation type: RWc (random walk with cycles), RWa (random walk acyclic, default), RWs (single source random walk).')
    parser.add_argument('--partial', action='store_true', default=False,
                        help='Test with partial prefixes: use a random portion of the path as prompt instead of just "A source target:" (default: False)')
    # New argument to handle data without task tags
    parser.add_argument('--no_task_tag', action='store_true', default=False,
                        help='Data files do not contain task identifiers (A, B, C, etc.). When enabled, task tokens will not be considered in data processing. This should match the setting used during data generation.')
    parser.add_argument('--num_labels', type=int, default=10,
                        help='Number of distinct node labels (default: 10). Must match data generation.')
    parser.add_argument('--PostGRU', action='store_true', default=False,
                        help='Load PostGRU checkpoint (adds _PGR suffix to checkpoint filename)')
    parser.add_argument('--NLS', action='store_true', default=False,
                        help='Load NLS checkpoint (adds _NLS suffix to checkpoint filename)')
    parser.add_argument('--DyadicAttn', action='store_true', default=False,
                        help='Load DyadicAttn checkpoint (adds _DA suffix to checkpoint filename)')
    parser.add_argument('--DyadicHybrid', action='store_true', default=False,
                        help='Load DyadicHybrid checkpoint (adds _DH suffix to checkpoint filename)')
    parser.add_argument('--ckpt_suffix', type=str, default='',
                        help='Extra checkpoint tag to select a specific variant trained by '
                             'train_taskC.py (e.g. "SA" loads ..._{tasks_tag}_SA_{train_label}.pt). '
                             'Default empty = baseline model.')
    # batch_size = 100 is for my laptop GPU with 4GB memory, for A100 GPU, you can set batch_size = 1000
    return parser.parse_args()


args = parse_args()
dataset = 'maze'
ckpt_iter = args.ckpt_iter
device = args.device
temperature = args.temperature
num_nodes = args.num_nodes
num_of_paths = args.num_of_paths
config = args.config
multitasks = args.multitasks
num_train_dataset = args.num_train_dataset
num_test_dataset = args.num_test_dataset
train_label = format_count(num_train_dataset)
test_dataset_label = format_count(num_test_dataset)
run_test_label = args.batch_size * args.num_iters
no_task_tag = args.no_task_tag  # Get the no_task_tag flag

tasks_str = args.tasks
tasks_tag = f"{tasks_str}_CL" if args.CL else tasks_str
# Parse path_type for filenames (RWc = cyclic, RWa = acyclic, RWs = single source)
allow_cycles = (args.path_type in ['RWc', 'RWs'])
path_type_tag = args.path_type
tasks_tag = f"{tasks_tag}_{path_type_tag}"
# Include num_labels in tags when non-default
if args.num_labels != 10:
    tasks_tag = f"{tasks_tag}_L{args.num_labels}"
# Add _NT_ tag to tasks_tag when no_task_tag is enabled
if args.no_task_tag:
    tasks_tag = f"{tasks_tag}_NT"
# data_tasks_tag: used for finding data/meta files (no _NL suffix)
data_tasks_tag = tasks_tag
# Add _NL tag to tasks_tag for transformer-nextlat (affects checkpoint and output naming only)
if args.model == 'transformer-nextlat':
    tasks_tag = f"{tasks_tag}_NL"
# Add _PGR tag to tasks_tag when PostGRU is enabled (affects checkpoint and output naming only)
if args.PostGRU:
    tasks_tag = f"{tasks_tag}_PGR"
# Add _NLS tag to tasks_tag when NLS is enabled (affects checkpoint and output naming only)
if args.NLS:
    tasks_tag = f"{tasks_tag}_NLS"
# Add _DA tag to tasks_tag when DyadicAttn is enabled
if args.DyadicAttn:
    tasks_tag = f"{tasks_tag}_DA"
if args.DyadicHybrid:
    tasks_tag = f"{tasks_tag}_DH"
# Append an arbitrary variant suffix (e.g. SA = state-aligned from train_taskC.py).
# Affects checkpoint loading and prediction output naming, not data/graph discovery.
if args.ckpt_suffix:
    tasks_tag = f"{tasks_tag}_{args.ckpt_suffix}"
# Graph tag includes path type to match generated graph files
graph_tag = f"{tasks_str}_CL" if args.CL else tasks_str
graph_tag = f"{graph_tag}_{path_type_tag}"
# Include num_labels in graph_tag when non-default
if args.num_labels != 10:
    graph_tag = f"{graph_tag}_L{args.num_labels}"
# Add _NT_ tag to graph_tag when no_task_tag is enabled
if args.no_task_tag:
    graph_tag = f"{graph_tag}_NT"

data_path = f'data/{dataset}/{num_nodes}'


def pick_first_existing_meta(candidates):
    for path in candidates:
        if os.path.exists(path):
            return path
    return candidates[0]


meta_path = pick_first_existing_meta([
    f'{data_path}/meta_{data_tasks_tag}.pkl',
    f'{data_path}/meta_{tasks_str}.pkl',
    f'{data_path}/meta.pkl',
])

print(f"Loading meta from {meta_path}...")
with open(meta_path, 'rb') as f:
    meta = pickle.load(f)

stoi, itos = meta['stoi'], meta['itos']
max_new_tokens = meta['block_size']
top_k = len(itos)
simple_format = meta['simple_format']

# Check if metadata contains no_task_tag flag and use it if available
if 'no_task_tag' in meta:
    meta_no_task_tag = meta['no_task_tag']

nt_suffix = '_NT' if no_task_tag else ''
out_dir = f'out/{args.model.replace("-", "_")}/{dataset}_{config}_{num_nodes}{nt_suffix}/'


def pick_first_existing(candidates):
    for path in candidates:
        if os.path.exists(path):
            return path
    return candidates[0]


if multitasks:
    candidate_ckpts = [
        os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze_{tasks_tag}_{train_label}.pt'),
        os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze_{tasks_tag}_{num_train_dataset}.pt'),
        os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze_{tasks_str}_{train_label}.pt'),
        os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze_{tasks_str}_{num_train_dataset}.pt'),
        os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze_{num_of_paths}.pt'),
        os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze.pt'),
    ]
    ckpt_path = pick_first_existing(candidate_ckpts)
else:
    if (num_of_paths == 0):
        ckpt_path = os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze.pt')
    else:
        ckpt_path = os.path.join(out_dir, f'{ckpt_iter}_ckpt_maze_{num_of_paths}.pt')
checkpoint = torch.load(ckpt_path, map_location=device, weights_only=False)
model_args = checkpoint['model_args']
ckpt_model_type = checkpoint.get('model_type', args.model)
if ckpt_model_type == 'mamba':
    model = Mamba(MambaConfig(**model_args))
elif ckpt_model_type == 'mamba2':
    model = Mamba2(Mamba2Config(**model_args))
elif ckpt_model_type == 'gated-deltanet':
    model = GatedDeltaNet(GatedDeltaNetConfig(**model_args))
elif ckpt_model_type == 'gru':
    model = GRU(GRUConfig(**model_args))
elif ckpt_model_type == 'transformer-nextlat':
    if args.local:
        model_args['use_flash'] = False  # Override for local GPU compatibility
    model = TransformerNextLat(TransformerNextLatConfig(**model_args))
elif ckpt_model_type == 'transformer-rope':
    if args.local:
        model_args['use_flash'] = False  # Override for local GPU compatibility
    model = GPTRoPE(GPTRoPEConfig(**model_args))
else:
    if args.local:
        model_args['use_flash'] = False  # Override for local GPU compatibility
    gptconf = GPTConfig(**model_args)
    model = GPT(gptconf)
state_dict = checkpoint['model']
unwanted_prefix = '_orig_mod.'
for k, v in list(state_dict.items()):
    if k.startswith(unwanted_prefix):
        state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
model.load_state_dict(state_dict)

model.eval()
model.to(device)

graph_file = args.graph_file
if graph_file is not None:
    maze_graph_path = graph_file if os.path.isabs(graph_file) else os.path.join(data_path, graph_file)
else:
    if multitasks:
        maze_graph_path = pick_first_existing([
            f'{data_path}/maze_graph_{graph_tag}.graphml',
            f'{data_path}/maze_graph_{tasks_str}.graphml',
            f'{data_path}/maze_graph.graphml',
        ])
    else:
        maze_graph_path = f'{data_path}/maze_graph.graphml'
maze_graph = nx.read_graphml(maze_graph_path)

# Discover node labels from graph (supports arbitrary label sets)
graph_node_labels = set(attrs.get('label', '?') for _, attrs in maze_graph.nodes(data=True))
print(f"Graph labels ({len(graph_node_labels)}): {sorted(graph_node_labels)[:10]}{'...' if len(graph_node_labels) > 10 else ''}")

# Load separate graph for Task B if specified
B_graph_file = args.B_graph_file
if B_graph_file is not None:
    B_graph_path = B_graph_file if os.path.isabs(B_graph_file) else os.path.join(data_path, B_graph_file)
    print(f"Loading Task B graph from {B_graph_path}...")
    maze_graph_B = nx.read_graphml(B_graph_path)
else:
    maze_graph_B = maze_graph  # Use the same graph for Task B

# Calculate grid size from number of nodes
n = int(math.sqrt(num_nodes))


def find_third_number_position(number_string):
    numbers = number_string.split()
    if no_task_tag:
        # In no_task_tag mode, there is no task ID, so source and target are at indices 0 and 1
        third_number_index = 2  # source + target
    else:
        # Check if first token is a task ID (A, B, C, D, E, F, G)
        if numbers[0] in ['A', 'B', 'C', 'D', 'E', 'F', 'G']:
            # Skip task ID, then get source and target nodes (indices 1 and 2)
            third_number_index = 3  # task_id + source + target
        else:
            # Old format without task ID
            third_number_index = 2  # source + target
    position = sum(len(num) for num in numbers[:third_number_index]) + third_number_index - 1
    return position


def encode(s):
    ss = s.split(" ")
    encoded_string = [stoi[ch] for ch in ss]
    return encoded_string


def decode(l):
    dec = ""
    for i in l:
        dec = dec + itos[i] + " "
    return dec[:-1]


def check_maze_path(G, gen_str, n, prefix_dir_count=0):
    """
    Check if a maze path in direction format is valid.
    Format: "task_id source_node target_node direction_sequence" or "source_node target_node direction_sequence"
    Task IDs: A, B, C, D, E, F, G (optional, for multi-task support)
    Directions: N (north/up), S (south/down), E (east/right), W (west/left)

    Args:
        G: The graph
        gen_str: The generated string to check
        n: Grid size
        prefix_dir_count: Number of prefix directions (for partial mode, step numbers in errors will be offset)

    Returns:
        '' if path is correct
        error message otherwise
    """
    tokens = [t for t in gen_str.split() if t != ':' and t != '>']

    # Check if first token is a task ID (only if not in no_task_tag mode)
    task_offset = 0
    if not no_task_tag and len(tokens) > 0 and tokens[0] in ['A', 'B', 'C', 'D', 'E', 'F', 'G']:
        task_offset = 1

    # Check basic syntax: need at least source and target (after task ID if present)
    if len(tokens) < 2 + task_offset:
        return 'syntax error'

    try:
        source = int(tokens[task_offset])
        target = int(tokens[task_offset + 1])
    except (ValueError, IndexError):
        return 'syntax error'

    # Validate node IDs
    if source < 0 or source >= num_nodes or target < 0 or target >= num_nodes:
        return 'syntax error'

    # Extract direction sequence (everything after task_id, source and target)
    directions = tokens[2 + task_offset:]

    # Start from source node
    current_node = source

    # Follow each direction
    for i, direction in enumerate(directions):
        if direction not in ['N', 'S', 'E', 'W']:
            return 'syntax error'

        # Calculate next node based on direction
        next_node = None
        if direction == 'N':  # North (up)
            next_node = current_node - n
        elif direction == 'S':  # South (down)
            next_node = current_node + n
        elif direction == 'E':  # East (right)
            next_node = current_node + 1
        elif direction == 'W':  # West (left)
            next_node = current_node - 1

        # Check if next_node is valid
        # For partial mode, report step number relative to suffix (subtract prefix_dir_count)
        suffix_step = i - prefix_dir_count
        if next_node is None or next_node < 0 or next_node >= num_nodes:
            if prefix_dir_count > 0 and suffix_step >= 0:
                return f'suffix_step {suffix_step} node {current_node} direction {direction} is illegal'
            else:
                return f'step {i} node {current_node} direction {direction} is illegal'

        # Check if edge exists in the graph
        if not G.has_edge(str(current_node), str(next_node)):
            if prefix_dir_count > 0 and suffix_step >= 0:
                return f'suffix_step {suffix_step} node {current_node} direction {direction} is illegal'
            else:
                return f'step {i} node {current_node} direction {direction} is illegal'

        # Move to next node
        current_node = next_node

    # Check if we reached the target
    if current_node != target:
        return 'incorrect target node'

    return ''


def check_turn_path(G, gen_str, n, cl_mode=False):
    """Validate a path expressed as relative turns (L/R/F/T).

    The agent starts facing East at the source node. Each token both turns
    and advances one step in the grid.

    When cl_mode is True, after each L or R turn token, there should be a
    node label token matching the current node (before moving).
    """

    tokens = [t for t in gen_str.split() if t != ':']

    task_offset = 0
    if not no_task_tag and len(tokens) > 0 and tokens[0] in TASK_TOKENS:
        task_offset = 1

    if len(tokens) < 2 + task_offset:
        return 'syntax error'

    try:
        source = int(tokens[task_offset])
        target = int(tokens[task_offset + 1])
    except (ValueError, IndexError):
        return 'syntax error'

    if source < 0 or source >= num_nodes or target < 0 or target >= num_nodes:
        return 'syntax error'

    actions = tokens[2 + task_offset:]
    orientation = 'E'  # starts facing east
    current_node = source

    left_of = {'N': 'W', 'W': 'S', 'S': 'E', 'E': 'N'}
    right_of = {v: k for k, v in left_of.items()}
    opposite_of = {'N': 'S', 'S': 'N', 'E': 'W', 'W': 'E'}
    delta = {'N': -n, 'S': n, 'E': 1, 'W': -1}
    node_labels = graph_node_labels

    action_idx = 0
    step = 0
    while action_idx < len(actions):
        action = actions[action_idx]
        if action not in ['L', 'R', 'F', 'T']:
            return 'syntax error'

        if action == 'F':
            next_orientation = orientation
        elif action == 'L':
            next_orientation = left_of[orientation]
        elif action == 'R':
            next_orientation = right_of[orientation]
        else:  # 'T'
            next_orientation = opposite_of[orientation]

        next_node = current_node + delta[next_orientation]
        if next_node < 0 or next_node >= num_nodes:
            return f'step {step} node {current_node} direction {action} is illegal'
        if not G.has_edge(str(current_node), str(next_node)):
            return f'step {step} node {current_node} direction {action} is illegal'

        # In CL mode, after L or R, expect a node label for the current node (checked after direction validity)
        if cl_mode and action in ['L', 'R']:
            if action_idx + 1 >= len(actions):
                return 'syntax error'  # missing label after L/R
            label_token = actions[action_idx + 1]
            if label_token not in node_labels:
                return 'syntax error'  # expected a node label
            expected_label = G.nodes[str(current_node)]['label']
            if label_token != expected_label:
                return f'step {step} incorrect label {label_token} (expected {expected_label})'
            action_idx += 1  # skip the label token

        orientation = next_orientation
        current_node = next_node
        action_idx += 1
        step += 1

    if current_node != target:
        return 'incorrect target node'

    return ''


# Task-aware correctness wrappers
TASK_TOKENS = {'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I'}


def check_correctness_taskA(G, gen_str, n, prefix_dir_count=0):
    # Task A shares the same validation as path finding.
    return check_maze_path(G, gen_str, n, prefix_dir_count=prefix_dir_count)


def check_correctness_taskB(G, gen_str, n, prompt_tokens=None):
    # Validate Task B outputs against the prompt (start + directions) and graph labels.
    if prompt_tokens is None:
        return 'syntax error'

    def strip_colon(seq):
        return [t for t in seq if t != ':']

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        # In no_task_tag mode, there is no 'B' task identifier
        if ':' not in tokens_raw:
            return 'syntax error'
        colon_idx = tokens_raw.index(':')
        prompt_part = tokens_raw[:colon_idx]
        answer_part = tokens_raw[colon_idx + 1:]
    else:
        # Original format with task ID
        if not tokens_raw or tokens_raw[0] != 'B':
            return 'syntax error'
        if ':' not in tokens_raw:
            return 'syntax error'
        colon_idx = tokens_raw.index(':')
        prompt_part = tokens_raw[:colon_idx]
        answer_part = tokens_raw[colon_idx + 1:]

    prompt_clean = strip_colon(prompt_tokens)
    if no_task_tag:
        if len(prompt_clean) < 1:
            return 'syntax error'
        try:
            start_node = int(prompt_clean[0])
        except Exception:
            return 'syntax error'
        directions = prompt_clean[1:]
    else:
        if len(prompt_clean) < 2 or prompt_clean[0] != 'B':
            return 'syntax error'
        try:
            start_node = int(prompt_clean[1])
        except Exception:
            return 'syntax error'
        directions = prompt_clean[2:]

    # Walk the prompt directions to find the true end node
    current = start_node
    for direction in directions:
        if direction not in ['N', 'S', 'E', 'W']:
            return 'syntax error'
        if direction == 'N':
            next_node = current - n
        elif direction == 'S':
            next_node = current + n
        elif direction == 'E':
            next_node = current + 1
        else:  # 'W'
            next_node = current - 1

        if next_node < 0 or next_node >= len(G.nodes):
            return 'syntax error'
        if not G.has_edge(str(current), str(next_node)):
            return 'syntax error'
        current = next_node

    true_end = current

    # Expect exactly: <target_label> <E> <S> <W> <N>
    if len(answer_part) != 5:
        return 'syntax error'

    pred_label = answer_part[0]
    true_label = G.nodes[str(true_end)]['label']
    if pred_label != true_label:
        return 'incorrect target node label'

    pred_neighbors = answer_part[1:5]

    neighbors_order = [(1, 'east'), (n, 'south'), (-1, 'west'), (-n, 'north')]
    for idx, (offset, dir_name) in enumerate(neighbors_order):
        neighbor_id = true_end + offset
        has_neighbor = G.has_edge(str(true_end), str(neighbor_id))
        true_neighbor_label = G.nodes[str(neighbor_id)]['label'] if has_neighbor else '/'
        if pred_neighbors[idx] != true_neighbor_label:
            return f'incorrect neighbor label {dir_name}'

    return ''


def check_correctness_taskC(G, gen_str, n, cl_mode=False):
    return check_turn_path(G, gen_str, n, cl_mode=cl_mode)


def check_correctness_taskD(G, gen_str, n, prompt_tokens=None):
    if prompt_tokens is None:
        return 'syntax error'

    prompt_clean = [t for t in prompt_tokens if t != ':']
    if no_task_tag:
        if len(prompt_clean) < 2:
            return 'syntax error'
        try:
            source = int(prompt_clean[0])
        except Exception:
            return 'syntax error'
        target_label = prompt_clean[1]
    else:
        if len(prompt_clean) < 3 or prompt_clean[0] != 'D':
            return 'syntax error'
        try:
            source = int(prompt_clean[1])
        except Exception:
            return 'syntax error'
        target_label = prompt_clean[2]

    node_labels = graph_node_labels
    if target_label not in node_labels:
        return 'syntax error'

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        if ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]
    else:
        if not tokens_raw or tokens_raw[0] != 'D' or ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]

    if not answer_part:
        return 'syntax error'

    current_node = source
    for i, direction in enumerate(answer_part):
        if direction not in ['N', 'S', 'E', 'W']:
            return 'syntax error'
        if direction == 'N':
            next_node = current_node - n
        elif direction == 'S':
            next_node = current_node + n
        elif direction == 'E':
            next_node = current_node + 1
        else:
            next_node = current_node - 1

        if next_node < 0 or next_node >= num_nodes:
            return f'step {i} node {current_node} direction {direction} is illegal'
        if not G.has_edge(str(current_node), str(next_node)):
            return f'step {i} node {current_node} direction {direction} is illegal'
        current_node = next_node

    end_label = G.nodes[str(current_node)]['label']
    if end_label != target_label:
        return 'incorrect target node'

    lengths = nx.single_source_shortest_path_length(G, str(source))
    min_dist = None
    for node_id, attrs in G.nodes(data=True):
        if attrs.get('label') != target_label:
            continue
        dist = lengths.get(str(node_id))
        if dist is None:
            continue
        if min_dist is None or dist < min_dist:
            min_dist = dist
    if min_dist is None or len(answer_part) != min_dist:
        return 'incorrect target node'

    return ''


def check_correctness_taskE(G, gen_str, n, prompt_tokens=None):
    if prompt_tokens is None:
        return "syntax error"

    prompt_clean = [t for t in prompt_tokens if t != ':']
    if no_task_tag:
        if len(prompt_clean) < 2:
            return "syntax error"
        try:
            source = int(prompt_clean[0]); target = int(prompt_clean[1])
        except Exception:
            return "syntax error"
    else:
        if len(prompt_clean) < 3 or prompt_clean[0] != 'E':
            return "syntax error"
        try:
            source = int(prompt_clean[1]); target = int(prompt_clean[2])
        except Exception:
            return "syntax error"

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        if ':' not in tokens_raw:
            return "syntax error"
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]
    else:
        if not tokens_raw or tokens_raw[0] != 'E' or ':' not in tokens_raw:
            return "syntax error"
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]

    if not answer_part:
        return "syntax error"

    if len(answer_part) % 2 != 0:
        return f"step {len(answer_part) - 1} syntax error"

    node_labels = graph_node_labels
    dirs = {'N', 'S', 'E', 'W'}

    runs = []
    for i in range(0, len(answer_part), 2):
        d = answer_part[i]
        lab = answer_part[i + 1]
        if d not in dirs:
            return f"step {i} syntax error"
        if lab not in node_labels:
            return f"step {i + 1} syntax error"

        if not runs or runs[-1][0] != d or runs[-1][1] != lab:
            runs.append([d, lab, 1, i])
        else:
            runs[-1][2] += 1

    try:
        total_nodes = G.number_of_nodes()
    except Exception:
        total_nodes = None

    current = source

    def step(node, direction):
        if direction == 'N': return node - n
        if direction == 'S': return node + n
        if direction == 'E': return node + 1
        return node - 1

    for run_idx, (direction, lab, k, start_tok_idx) in enumerate(runs):
        seen = 0
        step_cap = total_nodes + 5 if total_nodes is not None else 10000
        steps = 0

        while True:
            steps += 1
            if steps > step_cap:
                return f"step {start_tok_idx} run seems non-terminating"

            nxt = step(current, direction)

            if total_nodes is not None and (nxt < 0 or nxt >= total_nodes):
                return f"step {start_tok_idx} node {current} direction {direction} is illegal"

            if not G.has_edge(str(current), str(nxt)):
                return f"step {start_tok_idx} node {current} direction {direction} is illegal"

            current = nxt
            node_label = G.nodes[str(current)]['label']

            if node_label == lab:
                seen += 1
                if seen == k:
                    break

    if current != target:
        return "incorrect target node"

    return ""


def check_correctness_taskF(G, gen_str, n, prompt_tokens=None):
    if prompt_tokens is None:
        return 'syntax error'

    prompt_clean = [t for t in prompt_tokens if t != ':']
    if no_task_tag:
        if len(prompt_clean) < 1:
            return 'syntax error'
        start_label = prompt_clean[0]
        directions = prompt_clean[1:]
    else:
        if len(prompt_clean) < 3 or prompt_clean[0] != 'F':
            return 'syntax error'
        start_label = prompt_clean[1]
        directions = prompt_clean[2:]

    if start_label not in graph_node_labels:
        return 'syntax error'

    valid_target_labels = set()
    any_start = False

    for node_id, attrs in G.nodes(data=True):
        if attrs.get('label') != start_label:
            continue
        any_start = True
        current_node = int(node_id)
        for direction in directions:
            if direction not in ['N', 'S', 'E', 'W']:
                return 'syntax error'
            if direction == 'N':
                next_node = current_node - n
            elif direction == 'S':
                next_node = current_node + n
            elif direction == 'E':
                next_node = current_node + 1
            else:
                next_node = current_node - 1
            if next_node < 0 or next_node >= num_nodes:
                current_node = None
                break
            if not G.has_edge(str(current_node), str(next_node)):
                current_node = None
                break
            current_node = next_node
        if current_node is not None:
            valid_target_labels.add(G.nodes[str(current_node)]['label'])

    if not any_start or not valid_target_labels:
        return 'syntax error'

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        if ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]
    else:
        if not tokens_raw or tokens_raw[0] != 'F' or ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]

    if len(answer_part) != 1:
        return 'syntax error'

    if answer_part[0] not in valid_target_labels:
        return 'incorrect target label'

    return ''


def check_correctness_taskG(G, gen_str, n, prompt_tokens=None):
    if prompt_tokens is None:
        return 'syntax error'

    prompt_clean = [t for t in prompt_tokens if t != ':']
    if no_task_tag:
        if len(prompt_clean) < 4:
            return 'syntax error'
        try:
            source1 = int(prompt_clean[0])
            source2 = int(prompt_clean[1])
            target1 = int(prompt_clean[2])
            target2 = int(prompt_clean[3])
        except Exception:
            return 'syntax error'
    else:
        if len(prompt_clean) < 5 or prompt_clean[0] != 'G':
            return 'syntax error'
        try:
            source1 = int(prompt_clean[1])
            source2 = int(prompt_clean[2])
            target1 = int(prompt_clean[3])
            target2 = int(prompt_clean[4])
        except Exception:
            return 'syntax error'

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        if ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]
    else:
        if not tokens_raw or tokens_raw[0] != 'G' or ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]

    if len(answer_part) < 3:
        return 'syntax error'

    try:
        chosen_source = int(answer_part[0])
        chosen_target = int(answer_part[1])
    except Exception:
        return 'syntax error'

    if (chosen_source, chosen_target) not in [(source1, target1), (source2, target2)]:
        return 'syntax error'

    if not nx.has_path(G, str(chosen_source), str(chosen_target)):
        return 'incorrect target node'

    directions = answer_part[2:]
    current_node = chosen_source
    for i, direction in enumerate(directions):
        if direction not in ['N', 'S', 'E', 'W']:
            return 'syntax error'
        if direction == 'N':
            next_node = current_node - n
        elif direction == 'S':
            next_node = current_node + n
        elif direction == 'E':
            next_node = current_node + 1
        else:
            next_node = current_node - 1
        if next_node < 0 or next_node >= num_nodes:
            return f'step {i} node {current_node} direction {direction} is illegal'
        if not G.has_edge(str(current_node), str(next_node)):
            return f'step {i} node {current_node} direction {direction} is illegal'
        current_node = next_node

    if current_node != chosen_target:
        return 'incorrect target node'

    return ''


def check_correctness_taskH(G, gen_str, n, prompt_tokens=None):
    """Validate Task H: relative clockwise-index path encoding.

    The walker starts facing East. Each answer token is a 1-based index
    into the feasible edges enumerated clockwise starting from the first
    direction after the current facing direction.
    """
    if prompt_tokens is None:
        return 'syntax error'

    prompt_clean = [t for t in prompt_tokens if t != ':']
    if no_task_tag:
        if len(prompt_clean) < 2:
            return 'syntax error'
        try:
            source = int(prompt_clean[0])
            target = int(prompt_clean[1])
        except Exception:
            return 'syntax error'
    else:
        if len(prompt_clean) < 3 or prompt_clean[0] != 'H':
            return 'syntax error'
        try:
            source = int(prompt_clean[1])
            target = int(prompt_clean[2])
        except Exception:
            return 'syntax error'

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        if ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]
    else:
        if not tokens_raw or tokens_raw[0] != 'H' or ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]

    if not answer_part:
        return 'syntax error'

    CLOCKWISE_SCAN = {
        'N': ['N', 'E', 'S', 'W'],
        'E': ['E', 'S', 'W', 'N'],
        'S': ['S', 'W', 'N', 'E'],
        'W': ['W', 'N', 'E', 'S'],
    }
    DELTA = {'N': -n, 'S': n, 'E': 1, 'W': -1}

    facing = 'E'
    current = source

    for step_idx, token in enumerate(answer_part):
        try:
            idx = int(token)
        except ValueError:
            return f'step {step_idx} syntax error'

        if idx < 1 or idx > 4:
            return f'step {step_idx} invalid index {idx}'

        # Enumerate feasible directions clockwise from facing
        scan_order = CLOCKWISE_SCAN[facing]
        feasible = []
        for d in scan_order:
            neighbor = current + DELTA[d]
            if 0 <= neighbor < num_nodes and G.has_edge(str(current), str(neighbor)):
                feasible.append(d)

        if not feasible:
            return f'step {step_idx} no feasible edges from node {current}'

        if idx > len(feasible):
            return f'step {step_idx} index {idx} exceeds feasible count {len(feasible)}'

        direction = feasible[idx - 1]
        next_node = current + DELTA[direction]

        facing = direction
        current = next_node

    if current != target:
        return 'incorrect target node'

    return ''


def check_correctness_taskI(G, gen_str, n, prompt_tokens=None):
    """Validate Task I: absolute clockwise-index path encoding (fixed North).

    Like Task H but feasible edges are always enumerated clockwise from a
    FIXED North reference (N, E, S, W); the walker does not track facing.
    Each answer token is the 1-based index into the node's feasible edges
    in this fixed order.
    """
    if prompt_tokens is None:
        return 'syntax error'

    prompt_clean = [t for t in prompt_tokens if t != ':']
    if no_task_tag:
        if len(prompt_clean) < 2:
            return 'syntax error'
        try:
            source = int(prompt_clean[0])
            target = int(prompt_clean[1])
        except Exception:
            return 'syntax error'
    else:
        if len(prompt_clean) < 3 or prompt_clean[0] != 'I':
            return 'syntax error'
        try:
            source = int(prompt_clean[1])
            target = int(prompt_clean[2])
        except Exception:
            return 'syntax error'

    tokens_raw = gen_str.strip().split()
    if no_task_tag:
        if ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]
    else:
        if not tokens_raw or tokens_raw[0] != 'I' or ':' not in tokens_raw:
            return 'syntax error'
        answer_part = tokens_raw[tokens_raw.index(':') + 1:]

    if not answer_part:
        return 'syntax error'

    FIXED_SCAN = ['N', 'E', 'S', 'W']
    DELTA = {'N': -n, 'S': n, 'E': 1, 'W': -1}

    current = source

    for step_idx, token in enumerate(answer_part):
        try:
            idx = int(token)
        except ValueError:
            return f'step {step_idx} syntax error'

        if idx < 1 or idx > 4:
            return f'step {step_idx} invalid index {idx}'

        # Enumerate feasible directions clockwise from a fixed North reference
        feasible = []
        for d in FIXED_SCAN:
            neighbor = current + DELTA[d]
            if 0 <= neighbor < num_nodes and G.has_edge(str(current), str(neighbor)):
                feasible.append(d)

        if not feasible:
            return f'step {step_idx} no feasible edges from node {current}'

        if idx > len(feasible):
            return f'step {step_idx} index {idx} exceeds feasible count {len(feasible)}'

        direction = feasible[idx - 1]
        current = current + DELTA[direction]

    if current != target:
        return 'incorrect target node'

    return ''


def check_path_unreachable(G, gen_str, gt):
    path = re.findall(r'\d+|x', gen_str)
    if 'x' in path and len(path) < 4:
        return 0 if 'x' in gt else 1

    if 'x' in gt and 'x' not in gen_str:
        return 1

    return check_maze_path(G, gen_str, n)


typedata = 'test'
typedata_candidates = (
    [
        os.path.join(data_path, f'test_{data_tasks_tag}_{test_dataset_label}.txt'),
        os.path.join(data_path, f'test_{data_tasks_tag}_{num_test_dataset}.txt'),
        os.path.join(data_path, f'test_{data_tasks_tag}_{run_test_label}.txt'),
        os.path.join(data_path, f'test_{tasks_str}_{test_dataset_label}.txt'),
        os.path.join(data_path, f'test_{tasks_str}_{num_test_dataset}.txt'),
    ]
    if multitasks else [os.path.join(data_path, f'{typedata}.txt')]
)
typedata_path = pick_first_existing(typedata_candidates)
f = open(typedata_path, encoding='gbk')
texts = []
encode_texts = []
ground_truth = []
full_lines = []  # Store full lines for partial prefix mode

for line in f:
    line = line.strip()
    if not line:
        continue
    full_lines.append(line)  # Store full line for partial mode
    if multitasks:
        texts.append(line.split(':')[0] + ':')
        encode_texts.append(encode(line.split(':')[0] + ':'))
    else:
        pos = find_third_number_position(line)
        if (line[:pos] != ''):
            texts.append(line[:pos])
            encode_texts.append(encode(line[:pos]))

    ground_truth.append(line)

ground_truth = np.array(ground_truth)
# Convert to torch tensors without padding to avoid training distribution mismatch
encode_texts = [torch.tensor(seq, dtype=torch.long).to(device) for seq in encode_texts]

# NOTE: Padding is NOT used due to variable-length prompts causing model to learn [PAD] tokens.
# During mixed task training (A+B), Task B prompts are much longer than Task A, causing heavy
# padding of Task A prompts. The model then learns to generate [PAD] as a valid output token.
# [Wei: the above is the comment from copilot. I guess technically it is not learning to generate [PAD], it is just that
#  with [PAD] in the prompts, the model never learns how to generate proper output with [PAD] input, causing it to generate
#  wrong outputs.]
# Instead, we process prompts individually without padding to match training distribution.
# Commented code below shows the original padding approach:
#
# encode_texts = pad_sequence(
#     [torch.tensor(seq, dtype=torch.long) for seq in encode_texts],
#     batch_first=True,
#     padding_value=0
# ).to(device)

from tqdm import tqdm
import random

batch_size = args.batch_size

# Add temperature suffix when temperature is not default (1.0)
temp_suffix = f'_t{temperature}' if temperature != 1.0 else ''

# Modify filename for partial mode
if args.partial:
    pred_filename = (
        f'pred_{typedata}_{tasks_tag}_{ckpt_iter}_{run_test_label}{temp_suffix}_partial.txt'
        if multitasks else f'pred_{typedata}_{ckpt_iter}_{num_of_paths}{temp_suffix}_partial.txt'
    )
else:
    pred_filename = (
        f'pred_{typedata}_{tasks_tag}_{ckpt_iter}_{run_test_label}{temp_suffix}.txt'
        if multitasks else f'pred_{typedata}_{ckpt_iter}_{num_of_paths}{temp_suffix}.txt'
    )

with open(out_dir + pred_filename, 'w') as f:
    pass

wrong = 0
total = 0
partial_prefix_lengths = []  # Track prefix lengths for partial mode statistics
j = 1

for i in tqdm(range(args.num_iters), desc="Testing batches"):
    ix = torch.randint(len(encode_texts), (batch_size,))

    if args.partial:
        # For partial mode, create random prefixes from full lines
        x_list = []
        partial_texts = []  # Store the partial prefix text for this batch
        prefix_dir_counts = []  # Track number of direction tokens in each prefix
        for idx in ix:
            full_line = full_lines[idx]
            # Split into prompt and path parts
            if ':' in full_line:
                prompt_part = full_line.split(':')[0] + ':'  # "A source target:" or "source target:"
                path_part = full_line.split(':')[1].strip()  # directions after ":"
                path_tokens = path_part.split()

                if len(path_tokens) > 0:
                    # Random prefix length: at least 1 token, up to all but 1 token
                    max_prefix_len = max(1, len(path_tokens) - 1)
                    prefix_len = random.randint(1, max_prefix_len)
                    partial_prefix_lengths.append(prefix_len)
                    prefix_dir_counts.append(prefix_len)  # Track prefix direction count

                    # Create partial prefix: prompt + some path tokens
                    partial_prefix = prompt_part + ' ' + ' '.join(path_tokens[:prefix_len])
                    partial_texts.append(partial_prefix)
                    x_list.append(torch.tensor(encode(partial_prefix), dtype=torch.long).to(device).unsqueeze(0))
                else:
                    # No path tokens, use original prompt
                    partial_texts.append(prompt_part)
                    prefix_dir_counts.append(0)
                    x_list.append(encode_texts[idx].unsqueeze(0))
            else:
                # No colon, use original encoding
                partial_texts.append(texts[idx])
                prefix_dir_counts.append(0)
                x_list.append(encode_texts[idx].unsqueeze(0))
    else:
        x_list = [encode_texts[idx].unsqueeze(0) for idx in ix]  # Convert each to batch size 1

    x_gt = ground_truth[ix]

    with torch.no_grad():
        y_pred_list = []
        confidence_list = []  # Store confidence scores for each sample
        top3_tokens_list = []  # Store top-3 token indices for each sample
        top3_probs_list = []  # Store top-3 probabilities for each sample

        # Group samples in x_list by prompt length so each group can be
        # generated in a single batched forward (huge speedup, especially
        # with NLS where per-step CUDA-launch overhead dominates).
        # Preserve original order via idx tracking.
        from collections import defaultdict
        groups = defaultdict(list)  # prompt_len -> list of (orig_idx, x_tensor)
        for orig_i, x in enumerate(x_list):
            groups[x.size(1)].append((orig_i, x))

        # Pre-allocate result slots so we can write back in original order.
        y_pred_list = [None] * len(x_list)
        confidence_list = [None] * len(x_list)
        top3_tokens_list = [None] * len(x_list)
        top3_probs_list = [None] * len(x_list)

        for prompt_len, items in groups.items():
            xb = torch.cat([x for _, x in items], dim=0)  # (B, prompt_len)
            yb, conf_b, t3t_b, t3p_b = model.generate(
                xb, max_new_tokens, temperature=temperature, top_k=top_k, return_confidence=True)
            B = xb.size(0)
            if B == 1:
                # generate() returns flat per-time lists for B=1.
                orig_i = items[0][0]
                y_pred_list[orig_i] = decode(yb[0].tolist()).split('\n')[0]
                confidence_list[orig_i] = conf_b
                top3_tokens_list[orig_i] = t3t_b
                top3_probs_list[orig_i] = t3p_b
            else:
                for k, (orig_i, _) in enumerate(items):
                    y_pred_list[orig_i] = decode(yb[k].tolist()).split('\n')[0]
                    confidence_list[orig_i] = conf_b[k]
                    top3_tokens_list[orig_i] = t3t_b[k]
                    top3_probs_list[orig_i] = t3p_b[k]

    y_pred = y_pred_list

    batch_wrong = 0
    with open(out_dir + pred_filename, 'a') as f:
        for t, item in enumerate(y_pred):
            total += 1
            tokens = item.split()

            if no_task_tag:
                original_prompt = texts[ix[t]].split() if not args.partial else partial_texts[t].split()
                if ':' in item:
                    colon_idx_char = item.index(':')
                    answer_part = item[colon_idx_char + 1:].strip()
                    answer_tokens = answer_part.split()

                    if len(original_prompt) >= 2 and original_prompt[0].isdigit() and original_prompt[1].isdigit():
                        if len(answer_tokens) >= 2 and answer_tokens[1] in graph_node_labels:
                            task_id = 'E'
                        elif any(tok in ['L', 'R', 'F', 'T'] for tok in answer_tokens):
                            task_id = 'C'
                        else:
                            task_id = 'A'
                    elif len(original_prompt) >= 2 and original_prompt[0].isdigit():
                        has_directions_in_prompt = any(tok in ['N', 'S', 'E', 'W'] for tok in original_prompt[1:])
                        task_id = 'B' if (has_directions_in_prompt and len(answer_tokens) == 5) else None
                    elif len(original_prompt) >= 2 and original_prompt[0].isdigit() and \
                            original_prompt[1] in graph_node_labels:
                        task_id = 'D'
                    elif len(original_prompt) >= 2 and original_prompt[0] in graph_node_labels:
                        task_id = 'F'
                    elif len(original_prompt) >= 4 and all(token.isdigit() for token in original_prompt[:4]):
                        task_id = 'G'
                    else:
                        task_id = None
                else:
                    task_id = None
            else:
                task_id = tokens[0] if len(tokens) > 0 and tokens[0] in TASK_TOKENS else None

            prompt_tokens = texts[ix[t]].split() if not args.partial else partial_texts[t].split()
            pdc = prefix_dir_counts[t] if args.partial else 0
            output_task_label = f"({task_id}) " if (no_task_tag and task_id) else ""

            if task_id == 'A' or (task_id is None and not multitasks):
                symbol = check_correctness_taskA(maze_graph, item, n, prefix_dir_count=pdc)
            elif task_id == 'B':
                symbol = check_correctness_taskB(maze_graph_B, item, n, prompt_tokens=prompt_tokens)
            elif task_id == 'C':
                symbol = check_correctness_taskC(maze_graph, item, n, cl_mode=args.CL)
            elif task_id == 'D':
                symbol = check_correctness_taskD(maze_graph, item, n, prompt_tokens=prompt_tokens)
            elif task_id == 'E':
                symbol = check_correctness_taskE(maze_graph, item, n, prompt_tokens=prompt_tokens)
            elif task_id == 'F':
                symbol = check_correctness_taskF(maze_graph, item, n, prompt_tokens=prompt_tokens)
            elif task_id == 'G':
                symbol = check_correctness_taskG(maze_graph, item, n, prompt_tokens=prompt_tokens)
            elif task_id == 'H':
                symbol = check_correctness_taskH(maze_graph, item, n, prompt_tokens=prompt_tokens)
            elif task_id == 'I':
                symbol = check_correctness_taskI(maze_graph, item, n, prompt_tokens=prompt_tokens)
            else:
                symbol = check_maze_path(maze_graph, item, n, prefix_dir_count=pdc)

            if (symbol != ""):
                wrong += 1
                batch_wrong += 1

            error_confidence_info = ""
            if symbol != "" and t < len(confidence_list):
                error_pos = None

                # 1. Try to find step number from error message (Illegal direction/label)
                step_match = re.search(r'(?:step|suffix_step|run)\s+(\d+)', symbol)
                if step_match:
                    error_pos = int(step_match.group(1))

                # 2. Handle "incorrect target node" or "syntax error"
                elif symbol == 'incorrect target node' or 'syntax error' in symbol:
                    gen_tokens = item.split()
                    if ':' in gen_tokens:
                        # Error is usually at the end of the generated sequence (the \n or the token after colon)
                        error_pos = len(gen_tokens) - gen_tokens.index(':') - 1

                # 3. Handle Task B/F specific label errors
                elif 'incorrect' in symbol and 'label' in symbol:
                    gen_tokens = item.split()
                    if ':' in gen_tokens:
                        # For label tasks, errors often happen at the first or specific token of the answer
                        error_pos = 0  # Default to the first token of the answer

                # If we found a position, extract confidence data
                if error_pos is not None and error_pos < len(confidence_list[t]):
                    error_conf = confidence_list[t][error_pos]
                    top3_tok = top3_tokens_list[t][error_pos]
                    top3_prob = top3_probs_list[t][error_pos]
                    top3_strs = [itos[tok] if tok < len(itos) else f"<{tok}>" for tok in top3_tok]

                    # Determine mistake type based on gap between top1 and the chosen token
                    if (top3_prob[0] - error_conf) > 0.4:
                        mistake_type = "LOW-CONF"
                    else:
                        mistake_type = "HIGH-CONF"

                    error_confidence_info = f" [{mistake_type}: conf={error_conf:.4f}, top3={top3_strs}, probs=[{top3_prob[0]:.4f},{top3_prob[1]:.4f},{top3_prob[2]:.4f}]]"

            if args.partial:
                gen_tokens = item.split()
                if ':' in gen_tokens:
                    colon_idx = gen_tokens.index(':')
                    marker_pos = colon_idx + 1 + pdc
                    if marker_pos <= len(gen_tokens):
                        gen_tokens.insert(marker_pos, '>')
                    marked_item = ' '.join(gen_tokens)
                else:
                    marked_item = item
                f.write(output_task_label + marked_item + " " + symbol + error_confidence_info + '\n')
            else:
                f.write(output_task_label + item + " " + symbol + error_confidence_info + '\n')

    # Calculate and display accuracy for this batch
    batch_correct = batch_size - batch_wrong
    batch_accuracy = 100.0 * batch_correct / batch_size
    # Update tqdm description with real-time accuracy
    print(f"Batch {j}/{args.num_iters}: Accuracy = {batch_accuracy:.2f}% ({batch_correct}/{batch_size})")
    j = j + 1

    # Update overall accuracy in tqdm progress bar
    overall_accuracy = 100.0 * (total - wrong) / total if total > 0 else 0.0

# Final accuracy calculation
overall_accuracy = 100.0 * (total - wrong) / total if total > 0 else 0.0
print(f"\nTotal predictions: {total}")
print(f"Correct predictions: {total - wrong}")
print(f"Wrong predictions: {wrong}")
print(f"Overall accuracy: {overall_accuracy:.2f}%")

# Print partial mode statistics
if args.partial and partial_prefix_lengths:
    avg_prefix_len = sum(partial_prefix_lengths) / len(partial_prefix_lengths)
    print(f"\nPartial prefix mode statistics:")
    print(f"  Average prefix length: {avg_prefix_len:.2f} tokens")
    print(f"  Min prefix length: {min(partial_prefix_lengths)}")
    print(f"  Max prefix length: {max(partial_prefix_lengths)}")

# Automatically run analyze_maze.py with the same arguments
import subprocess
import sys

print("\n" + "=" * 60)
print("Running analyze_maze.py...")
print("=" * 60)

analyze_cmd = [
    sys.executable, 'analyze_maze.py',
    '--ckpt_iter', str(args.ckpt_iter),
    '--model', args.model,
    '--config', config,
    '--dataset', dataset,
    '--num_nodes', str(args.num_nodes),
    '--num_of_paths', str(args.num_of_paths),
    '--num_train_dataset', str(num_train_dataset),
    '--num_test_dataset', str(num_test_dataset),
    '--tasks', tasks_str,
    '--batch_size', str(args.batch_size),
    '--num_iters', str(args.num_iters),
    '--path_type', args.path_type,
]

if args.multitasks:
    analyze_cmd.append('--multitasks')
else:
    analyze_cmd.append('--no-multitasks')

if args.CL:
    analyze_cmd.append('--CL')

if args.partial:
    analyze_cmd.append('--partial')

if args.no_task_tag:
    analyze_cmd.append('--no_task_tag')

if args.PostGRU:
    analyze_cmd.append('--PostGRU')

if args.NLS:
    analyze_cmd.append('--NLS')

if args.DyadicAttn:
    analyze_cmd.append('--DyadicAttn')

if args.DyadicHybrid:
    analyze_cmd.append('--DyadicHybrid')

if args.temperature != 1.0:
    analyze_cmd.extend(['--temperature', str(args.temperature)])

subprocess.run(analyze_cmd)