readCtrl_lambda / code /interface /old /label_thresold_v4.py
mshahidul
Initial commit of readCtrl code without large models
030876e
Raw
History Blame Contribute Delete
11.5 kB
import gradio as gr
import json
import os
import random
from datetime import datetime
# --- Configuration & Folder Setup ---
DATA_PATH = '/home/mshahidul/readctrl/data/extracting_subclaim/extracted_subclaims_syn_data_with_gs_summary_en.json'
KEY_DATA_PATH = '/home/mshahidul/readctrl/data/key_subclaims_testing/key_subclaims.json'
BASE_SAVE_DIR = '/home/mshahidul/readctrl/data/thresold_finding/'
# 1. Create folder based on date+hour of app start
session_folder_name = datetime.now().strftime("%Y-%m-%d_%Hh")
SESSION_PATH = os.path.join(BASE_SAVE_DIR, session_folder_name)
os.makedirs(SESSION_PATH, exist_ok=True)
# --- Data Loading ---
with open(DATA_PATH, 'r') as f:
data = json.load(f)
NUM_SAMPLES= 10
random.seed(42)
all_possible_indices = list(range(len(data)))
shuffled_indices = random.sample(all_possible_indices, min(NUM_SAMPLES, len(data)))
with open(KEY_DATA_PATH, 'r') as f:
key_data = json.load(f)
key_lookup = {item['index']: item['llm_output'] for item in key_data}
# --- Logic Functions ---
def get_key_indices(index, source_type):
if index not in key_lookup:
return []
key_field = 'key_source_text_subclaims' if source_type == "Full Original Text" else 'key_gold_summary_subclaims'
id_key = "source_subclaim_id" if source_type == "Full Original Text" else "gold_subclaim_id"
key_items = key_lookup[index].get(key_field, [])
indices = []
for item in key_items:
raw_id = item.get(id_key, "")
try:
idx = int(raw_id.split('-')[-1])
indices.append(idx)
except (ValueError, IndexError):
continue
return indices
def load_example(progress_index):
# Check if we've reached the end of our fixed sample size
if progress_index >= len(shuffled_indices):
return [
gr.update(value="### 🎉 Session Complete!"),
gr.update(value=f"You have finished your set of {NUM_SAMPLES} records."),
[], "0%", "0%", "0%", gr.update(choices=[], value=[]),
gr.update(choices=[], value=[]), gr.update(choices=[], value=[]), ""
]
# Get the actual index from our sample pool
actual_data_index = shuffled_indices[progress_index]
record = data[actual_data_index]
# Seed by actual_data_index for consistency
random.seed(actual_data_index)
source_type = random.choice(["Full Original Text", "Gold Summary"])
if source_type == "Full Original Text":
text_content, subclaims = record['fulltext'], record['fulltext_subclaims']
else:
text_content, subclaims = record['summary'], record['summary_subclaims']
source_info = f"### Instance: {progress_index + 1}/{len(shuffled_indices)} | Source: **{source_type}**"
key_indices = get_key_indices(actual_data_index, source_type)
pre_selected = [subclaims[idx] for idx in key_indices if 0 <= idx < len(subclaims)]
return [
source_info, text_content, subclaims, "0%", "0%", "0%",
gr.update(choices=subclaims, value=pre_selected),
gr.update(choices=subclaims, value=pre_selected),
gr.update(choices=subclaims, value=pre_selected),
""
]
def sync_from_low(low, inter, prof, total_list):
# Everything in Low must be in Intermediate and Proficient
new_inter = list(set(inter) | set(low))
new_prof = list(set(prof) | set(new_inter))
return update_ui_components(low, new_inter, new_prof, total_list)
def sync_from_inter(low, inter, prof, total_list):
# 1. Proficient must include everything in Intermediate
new_prof = list(set(prof) | set(inter))
# 2. Low can only contain items that are in Intermediate
new_low = list(set(low) & set(inter))
return update_ui_components(new_low, inter, new_prof, total_list)
def sync_from_prof(low, inter, prof, total_list):
# Intermediate and Low can only contain items that are in Proficient
new_inter = list(set(inter) & set(prof))
new_low = list(set(low) & set(prof))
return update_ui_components(new_low, new_inter, prof, total_list)
def update_ui_components(low, inter, prof, total_list):
"""Helper to calculate percentages and return updates for all groups"""
if not total_list:
return "0%", "0%", "0%", "", low, inter, prof
l_pct, i_pct, p_pct = (len(x)/len(total_list) * 100 for x in [low, inter, prof])
# Validation is now redundant because the code enforces it,
# but we can keep a success message.
msg = "✅ Hierarchy Enforced: Low ⊆ Intermediate ⊆ Proficient"
return (
f"{l_pct:.1f}%", f"{i_pct:.1f}%", f"{p_pct:.1f}%", msg,
gr.update(value=low), gr.update(value=inter), gr.update(value=prof)
)
def save_and_next(username, progress_index, source_info, low_sel, int_sel, prof_sel, subclaims):
"""
Saves the current annotation and moves to the next record in the random sample.
progress_index: The sequence number (0, 1, 2...) from the shuffled list.
shuffled_indices: This must be the global list generated at the top of your script.
"""
# 1. Validation: Ensure we haven't exceeded the sample size
if progress_index >= len(shuffled_indices):
return [progress_index] + load_example(progress_index)
# 2. Validation: Annotator Name
if not username or username.strip() == "":
gr.Warning("Action Required: Please enter your name before submitting!")
# Return current state to avoid losing work
return [progress_index, source_info, gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(),
gr.update(value=low_sel), gr.update(value=int_sel), gr.update(value=prof_sel),
"⚠️ **Error:** Please enter your name."]
# 3. Validation: Hierarchy Check (Low <= Intermediate <= Proficient)
if not (len(low_sel) <= len(int_sel) <= len(prof_sel)):
gr.Warning("DATA NOT SAVED! The selection does not follow the hierarchy: Low ≤ Intermediate ≤ Proficient.")
return [progress_index, source_info, gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(),
gr.update(value=low_sel), gr.update(value=int_sel), gr.update(value=prof_sel),
"⚠️ **Error:** Selection sequence is invalid. Please adjust before saving."]
# 4. Map progress to the actual data index from your JSON
actual_data_index = shuffled_indices[progress_index]
# 5. File System Management
try:
if not os.path.exists(SESSION_PATH):
os.makedirs(SESSION_PATH, exist_ok=True)
except Exception as e:
gr.Error(f"Critical Error: Could not create directory {SESSION_PATH}. Error: {e}")
return [progress_index] + load_example(progress_index)
# 6. Prepare Metadata and Filename
now = datetime.now()
timestamp_str = now.strftime("%Y%m%d_%H%M%S")
safe_username = "".join(x for x in username if x.isalnum())
# Use actual_data_index so you can easily match this file back to your master JSON
filename = f"recordID{actual_data_index}_seq{progress_index}_{safe_username}_{timestamp_str}.json"
file_path = os.path.join(SESSION_PATH, filename)
stype = "Full Original Text" if "Full Original Text" in source_info else "Gold Summary"
# 7. Construct Result Object
result = {
"annotator": username,
"timestamp": now.strftime("%Y-%m-%d %H:%M:%S"),
"progress_sequence": progress_index, # The order it was shown
"original_data_index": actual_data_index, # The real ID in the source JSON
"source_type": stype,
"annotations": {
"low": {
"count": len(low_sel),
"subclaims": low_sel,
"pct": len(low_sel)/len(subclaims) if subclaims else 0
},
"intermediate": {
"count": len(int_sel),
"subclaims": int_sel,
"pct": len(int_sel)/len(subclaims) if subclaims else 0
},
"proficient": {
"count": len(prof_sel),
"subclaims": prof_sel,
"pct": len(prof_sel)/len(subclaims) if subclaims else 0
}
}
}
# 8. Write to Disk
with open(file_path, 'w') as f:
json.dump(result, f, indent=4)
gr.Info(f"Success! Record {actual_data_index} saved (Item {progress_index + 1} of {len(shuffled_indices)}).")
# 9. Return the NEXT progress index and its data
return [progress_index + 1] + load_example(progress_index + 1)
# --- UI Definition ---
with gr.Blocks(theme=gr.themes.Soft(), title="Medical Literacy Annotation Tool") as demo:
index_state = gr.State(0)
subclaim_list_state = gr.State([])
try:
with open("/home/mshahidul/readctrl/code/interface/instructions", "r") as f:
instructions_text = f.read()
except:
instructions_text = "# Medical Annotation Task"
gr.Markdown(instructions_text)
with gr.Row():
with gr.Column(scale=1, variant="panel"):
user_input = gr.Textbox(label="Annotator Name", placeholder="e.g., mshahidul", interactive=True)
gr.HTML("<hr>")
source_display = gr.Markdown("### Initializing...")
text_viewer = gr.Textbox(label="Reference Text", interactive=False, lines=15)
with gr.Column(scale=2):
hierarchy_warning = gr.Markdown(value="", visible=True)
with gr.Row():
with gr.Column():
gr.Markdown("### 🟢 Low")
low_pct = gr.Label(label="Coverage", value="0%")
low_check = gr.CheckboxGroup(label="Subclaims", choices=[])
with gr.Column():
gr.Markdown("### 🟡 Intermediate")
int_pct = gr.Label(label="Coverage", value="0%")
int_check = gr.CheckboxGroup(label="Subclaims", choices=[])
with gr.Column():
gr.Markdown("### 🔴 Proficient")
prof_pct = gr.Label(label="Coverage", value="0%")
prof_check = gr.CheckboxGroup(label="Subclaims", choices=[])
submit_btn = gr.Button("Submit & Next Record", variant="primary", size="lg")
# Event Handlers
demo.load(load_example, [index_state], [source_display, text_viewer, subclaim_list_state, low_pct, int_pct, prof_pct, low_check, int_check, prof_check, hierarchy_warning])
# Event Handlers for Hierarchy Synchronization
low_check.input(
sync_from_low,
[low_check, int_check, prof_check, subclaim_list_state],
[low_pct, int_pct, prof_pct, hierarchy_warning, low_check, int_check, prof_check]
)
int_check.input(
sync_from_inter,
[low_check, int_check, prof_check, subclaim_list_state],
[low_pct, int_pct, prof_pct, hierarchy_warning, low_check, int_check, prof_check]
)
prof_check.input(
sync_from_prof,
[low_check, int_check, prof_check, subclaim_list_state],
[low_pct, int_pct, prof_pct, hierarchy_warning, low_check, int_check, prof_check]
)
submit_btn.click(save_and_next, [user_input, index_state, source_display, low_check, int_check, prof_check, subclaim_list_state], [index_state, source_display, text_viewer, subclaim_list_state, low_pct, int_pct, prof_pct, low_check, int_check, prof_check, hierarchy_warning])
if __name__ == "__main__":
demo.launch(share=True)