reading-study / app.py
BVendeville's picture
annotation app code-82b7d0ec0e7b
9029357 verified
Raw History Blame Contribute Delete
35.5 kB
"""Gradio annotation app for the human evaluation (HUMAN_EVALUATION.md §4). UI wiring only; logic is in core.py.
python app.py --data example_input.json [--config config.yaml] [--host 0.0.0.0] [--port 7860]
Every option has an env var (ANNOTATION_DATA, ANNOTATION_CONFIG, ANNOTATION_OUTPUT_DIR, ANNOTATION_LAYOUT,
ANNOTATION_HUB_REPO) because a Hugging Face Space starts the app with no arguments. With ANNOTATION_HUB_REPO set,
config, documents and answers live in that private dataset repo (hub.py).
"""
from __future__ import annotations
import argparse
import atexit
import html
import inspect
import os
import signal
import sys
import tempfile
from pathlib import Path
os.environ["GRADIO_ANALYTICS_ENABLED"] = "False" # before gradio is imported: no telemetry, no version check
import gradio as gr # noqa: E402
HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE))
import core as C # noqa: E402
import guide as G # noqa: E402
CSS = """
:root { color-scheme: light; } /* light theme only, see FRAME_FIT_JS */
/* Each text is one card: its name is a header bar INSIDE the card's border, so a name can only belong to the
text it frames, even when the cards are stacked. */
.ann-card { border: 2px solid var(--border-color-primary); border-radius: 8px; overflow: hidden; }
.ann-card-head { font-size: 1.15rem; font-weight: 700; padding: 0.45rem 1rem;
background: var(--background-fill-secondary); border-bottom: 1px solid var(--border-color-primary); }
.ann-card-source { border-color: var(--color-accent, #f97316); }
.ann-card-source .ann-card-head { background: var(--color-accent-soft, #fff7ed); }
.ann-text { max-height: 70vh; overflow-y: auto; padding: 0.5rem 1rem; line-height: 1.55; }
.ann-card-source .ann-text { max-height: 45vh; }
.ann-vertical { gap: 2rem !important; } /* clear space between stacked cards */
.ann-vertical .ann-card:not(.ann-card-source) .ann-text { max-height: none; } /* stacked texts grow to full length */
/* the participant's ID: hard to miss */
.ann-idwarn { background: #fff7e0; border: 2px solid #e0a800; border-radius: 8px; padding: 0.7rem 1rem; line-height: 1.6; }
.ann-idbox { background: #fff7e0; border: 2px solid #e0a800; border-radius: 10px; padding: 1rem 1.2rem;
font-size: 1.1rem; line-height: 1.6; margin-bottom: 0.5rem; }
.ann-id { font-family: ui-monospace, monospace; font-size: 1.6rem; font-weight: 700; padding: 0.1rem 0.5rem;
background: #fff; border: 1px dashed #b38600; border-radius: 6px; margin-left: 0.3rem; }
/* the questions: large, normal weight, only the key words bold (written in Markdown) */
.ann-qtext p { font-size: 1.25rem !important; font-weight: 400 !important; line-height: 1.45; margin: 0.9rem 0 0.2rem; }
.ann-qtext strong { font-weight: 700; }
/* instructions panel: its own tinted background and a coloured header bar, so it stands apart from the task */
#guide { background: #eef4fb !important; border: 2px solid #2f6db5 !important; border-radius: 10px !important; }
#guide > button.label-wrap { background: #2f6db5; color: #fff; font-weight: 700; font-size: 1.05rem;
padding: 0.65rem 1rem; margin: -1px; width: calc(100% + 2px);
border-radius: 8px 8px 0 0; }
#guide > button.label-wrap span { color: #fff; }
#guide .prose, #guide .md { background: transparent !important; }
/* instructions panel: real titles, not body-size bold lines */
#guide h3 { font-size: 1.45rem !important; font-weight: 700; margin: 1.4rem 0 0.5rem;
padding-bottom: 0.2rem; border-bottom: 1px solid var(--border-color-primary); }
#guide h4 { font-size: 1.15rem !important; font-weight: 700; margin: 1.1rem 0 0.35rem; }
#guide p, #guide li { line-height: 1.6; }
.ann-error { color: var(--error-text-color, #b91c1c); font-weight: 600; }
"""
GRADIO_6 = int(gr.__version__.split(".")[0]) >= 6
# Events are UI-only: no API endpoint a script could call to skip the screen order.
EVENT = {"api_visibility": "private"} if GRADIO_6 else {"show_api": False, "api_name": False}
def theme():
"""System fonts only: the default theme would fetch Google Fonts."""
return gr.themes.Default(font=["system-ui", "sans-serif"], font_mono=["ui-monospace", "monospace"])
# Inside the Hugging Face Space page the app sits in an iframe whose height gradio reports itself; gradio stops
# reporting after a few successive grows, so long screens got cut off (submit button out of reach). This keeps the
# frame as tall as the content: it measures the bottom of the last element (not the document, which can stretch to
# the frame and would loop) and asks iframe-resizer for that height whenever it changes.
# It also scrolls back to the top whenever a new screen is shown (the progress line changes): the submit button is at
# the bottom, so otherwise the next document opened on its last text and the questions, with "Document N" and the
# first texts out of view. It also keeps the page in the light theme whatever the reader's dark-mode setting.
FRAME_FIT_JS = """
<script>
(function () {
let last = 0, shown = '';
function fit() {
const pif = window.parentIFrame, end = document.getElementById('bottom_spacer');
if (!pif || !end) return;
const h = Math.ceil(end.getBoundingClientRect().bottom + window.scrollY + 24);
if (Math.abs(h - last) > 2) { last = h; pif.size(h); }
}
function toTop() {
// only the "Document N of M" part: while an event runs, gradio's status timer adds text to the element
const p = [...document.querySelectorAll('#s1_progress, #s2_progress')].find(e => e.offsetParent !== null);
const m = p ? p.textContent.match(/Document \\d+ of \\d+/) : null;
const key = m ? p.id + '|' + m[0] : null; // the id too: screens 1 and 2 of one document share the label
if (!key || key === shown) return;
shown = key;
window.scrollTo(0, 0);
const pif = window.parentIFrame;
if (pif && pif.scrollToOffset) pif.scrollToOffset(0, 0);
}
function light() {
// always the light theme: the custom colours assume it, and gradio's dark mode (from the system setting or
// the Space page) gave white text on white panels
for (const el of [document.documentElement, document.body]) if (el && el.classList.contains('dark')) el.classList.remove('dark');
}
light();
new MutationObserver(() => { light(); fit(); toTop(); }).observe(document.documentElement,
{subtree: true, childList: true, attributes: true});
window.addEventListener('resize', fit);
setInterval(fit, 1000);
})();
</script>
"""
def launch_kwargs(host=None, port=None) -> dict:
"""Keyword arguments for demo.launch(), keeping only those this gradio version accepts."""
kw = {"share": False, "ssr_mode": False, "server_name": host, "server_port": port,
"show_api": False, # gradio 5: no API page
"footer_links": [], # gradio 6: no footer (API docs, settings, run-history links)
"run_history": False, # gradio 6: do not keep past runs in the browser
"max_file_size": 1024, # the app takes no files: refuse anonymous uploads to the upload route
"pwa": False,
"head": FRAME_FIT_JS} # no installable-app manifest (on by default on a Space)
if GRADIO_6:
kw.update(theme=theme(), css=CSS)
accepted = inspect.signature(gr.Blocks.launch).parameters
return {k: v for k, v in kw.items() if k in accepted and v is not None}
def _card(title: str, text: str, source: bool = False) -> str:
body = "".join("<p>" + "<br>".join(html.escape(line) for line in block) + "</p>" for block in C.paragraphs(text))
return (f'<div class="ann-card{" ann-card-source" if source else ""}"><div class="ann-card-head">{title}</div>'
f'<div class="ann-text">{body}</div></div>')
def _id_box(code: str) -> str:
return ('<div class="ann-idbox">Your ID is <span class="ann-id">' + html.escape(code) + '</span>'
'<div>Write it down or take a screenshot <b>now</b>. You need it to come back to where you stopped, '
'and nobody can recover it for you.</div></div>')
def qmd(text: str, eid: str, visible: bool = True):
"""A question line: normal weight, large, with only its key words in bold (Markdown)."""
return gr.Markdown(text, elem_id=eid, elem_classes=["ann-qtext"], visible=visible)
def _err(msg: str | None) -> str:
return f'<div class="ann-error">{html.escape(msg)}</div>' if msg else ""
LAYOUTS = ("horizontal", "vertical")
TASKS = ("both", "simplicity", "validity") # both = screen 1 then screen 2; or one screen per Space
# Shown under the study title, so each Space says which part of the study it is and what to do.
TASK_INTRO = {
"both": "",
"simplicity": "### How easy is each text to follow?\n"
"For each document you read three simplifications of the same scientific text, written for "
"non-expert readers, and say which is easiest and which is hardest to follow. You do not see the "
"original text here.\n\nThe texts are in English: please take part only if you read English comfortably.",
"validity": "### Does each text match the original?\n"
"For each document you read the original scientific text and three simplifications of it, written "
"for non-expert readers, and tick the simplifications that leave out important information, "
"contradict the original, or state a finding more broadly than it supports.\n\nThe texts are in English: please take part only if you read English comfortably.",
}
def build_demo(docs, config, out_dir: str | Path, meta: dict, layout: str = "horizontal", on_saved=None,
on_login=None, remote_has=None, task: str = "both", split: bool = False) -> gr.Blocks:
"""``docs`` and ``config`` are the frozen objects from core.load_data / core.load_config; they are only read.
``layout``: the three texts side by side ("horizontal") or one above another ("vertical").
``on_saved(code)`` is called after a line is written (hub.AnswerSync.push, to copy it to the private repo).
``on_login(code)`` is called before the resume position is computed (hub.AnswerSync.pull).
``remote_has(code)`` says whether the private repo already has that ID (hub.AnswerSync.remote_has).
``task``: "both" (simplicity without the source, locked, then validity with it), "simplicity" (screen 1 only;
the source is never shown) or "validity" (screen 2 only: source and texts together).
``split``: each person gets one of the four blocks (a quarter of the documents) instead of all of them."""
if layout not in LAYOUTS:
raise ValueError(f"layout must be one of {LAYOUTS}, got {layout!r}")
if task not in TASKS:
raise ValueError(f"task must be one of {TASKS}, got {task!r}")
meta = {**meta, "layout": layout, "task": task}
def texts_box():
return gr.Row(equal_height=False) if layout == "horizontal" else gr.Column(elem_classes=["ann-vertical"])
n_docs = len(docs)
systems = list(docs[0]["simplifications"])
salt = str(config["study"].get("position_salt", ""))
study = config["study"]
style = {} if GRADIO_6 else {"theme": theme(), "css": CSS} # gradio 6 takes them in launch()
# delete_cache: anything left in gradio's temp dir (e.g. junk sent to the upload route) is removed hourly
with gr.Blocks(title=study.get("title", "Annotation"), analytics_enabled=False, delete_cache=(3600, 3600),
**style) as demo:
state = gr.State(None)
gr.Markdown(f"## {study.get('title', 'Annotation')}")
if TASK_INTRO[task]:
gr.Markdown(TASK_INTRO[task], elem_id="task_intro")
guide_md = study.get(f"guide_{task}") or G.GUIDE.get(task)
if guide_md: # always on the page, so it can be reopened at any time
with gr.Accordion("Instructions and example (click to open or close at any time)", open=False,
elem_id="guide"):
gr.Markdown(guide_md)
with gr.Column(visible=True) as g_start:
key_in = gr.Textbox(label="Access key", type="password", placeholder="the key you were sent",
visible=bool(config.get("access_key") or config.get("expert_access_key")
or config.get("block_keys")),
elem_id="key_in")
if task != "both":
gr.Markdown("Please read **Instructions and example** above before you start; you can open it again "
"at any time, on every page.")
gr.HTML('<div class="ann-idwarn"><b>First time?</b> Choose an ID (not your name). '
'<b>Write it down now, somewhere you will find it again</b> (on paper, in your phone, in a note): '
'it is the <b>only</b> way to come back to where you stopped. We cannot look it up or reset it '
'for you.<br><b>Coming back?</b> Enter the ID you chose last time.</div>')
code_in = gr.Textbox(label="Your ID", placeholder="3 to 30 letters, digits, - or _", elem_id="code_in")
with gr.Row():
new_btn = gr.Button("I'm new: start with this ID", variant="primary", elem_id="new_btn")
code_btn = gr.Button("I'm coming back: continue", variant="secondary", elem_id="code_btn")
start_err = gr.HTML("", elem_id="start_err")
with gr.Group(visible=False) as g_consent:
id_box = gr.HTML("", elem_id="id_box") # the participant's ID, shown big, before anything else
id_saved_cb = gr.Checkbox(label="I have saved my ID somewhere I will find it again.", value=False,
elem_id="id_saved_cb")
gr.Markdown(study.get(f"consent_markdown_{task}") or study["consent_markdown"]) # per-Space text
consent_cb = gr.Checkbox(label="I have read the above and agree to take part.", value=False, elem_id="consent_cb")
consent_btn = gr.Button("Continue", variant="primary", elem_id="consent_btn")
consent_err = gr.HTML("", elem_id="consent_err")
with gr.Column(visible=False) as g_s1:
s1_progress = gr.Markdown("", elem_id="s1_progress")
gr.Markdown("Read the three texts below, then answer the questions.")
with texts_box():
s1_texts = []
for L in C.LETTERS:
with gr.Column():
s1_texts.append(gr.HTML("", elem_id=f"s1_text_{L}"))
# each question is a Markdown line (only its key words in bold) above an unlabeled answer component
qmd("Which of these did you find **easiest to follow**?", "q_easiest")
easiest = gr.Radio(list(C.LETTERS), label=C.Q_EASIEST, show_label=False, value=None, elem_id="easiest")
qmd("Which did you find **hardest to follow**?", "q_hardest")
hardest = gr.Radio(list(C.LETTERS), label=C.Q_HARDEST, show_label=False, value=None, elem_id="hardest")
gap_bm_q = qmd("", "q_gap_bm", visible=False)
gap_bm = gr.Radio(list(C.GAPS), label="", show_label=False, value=None, visible=False, elem_id="gap_bm")
gap_mw_q = qmd("", "q_gap_mw", visible=False)
gap_mw = gr.Radio(list(C.GAPS), label="", show_label=False, value=None, visible=False, elem_id="gap_mw")
s1_err = gr.HTML("", elem_id="s1_err")
s1_btn = gr.Button("Submit and go to the next document" if task == "simplicity" else
"Submit and continue (answers cannot be changed afterwards)", variant="primary",
elem_id="s1_btn")
with gr.Column(visible=False) as g_s2:
s2_progress = gr.Markdown("", elem_id="s2_progress")
gr.Markdown("Compare each text with the **source** it was written from." if task == "validity" else
"Now compare each text with the **source** it was written from.")
s2_source = gr.HTML("", elem_id="s2_source") # stays empty until screen 1 is locked
with texts_box():
s2_texts = []
for L in C.LETTERS:
with gr.Column():
s2_texts.append(gr.HTML("", elem_id=f"s2_text_{L}"))
qmd("Does the simplification **leave out important information** from the source?", "q_omits")
omits = gr.CheckboxGroup(list(C.LETTERS), label=C.Q_OMITS, show_label=False, value=[], elem_id="omits")
qmd("Does the simplification **contradict the source**?", "q_contradicts")
contradicts = gr.CheckboxGroup(list(C.LETTERS), label=C.Q_CONTRADICTS, show_label=False, value=[],
elem_id="contradicts")
overgen_q = qmd("Does the simplification **state a finding more broadly than the source supports**? For "
"example dropping who it applies to, or under what conditions.", "q_overgen", visible=False)
overgen = gr.CheckboxGroup(list(C.LETTERS), label=C.Q_OVERGENERALISES, show_label=False, value=[],
visible=False, elem_id="overgen")
gr.Markdown("There may well be no error: leave a question unticked if none of the texts applies.")
s2_err = gr.HTML("", elem_id="s2_err")
s2_btn = gr.Button("Submit and go to the next document", variant="primary", elem_id="s2_btn")
with gr.Group(visible=False) as g_done:
done_md = gr.Markdown("", elem_id="done_md")
# end marker: FRAME_FIT_JS sizes the Space page's frame to its bottom; a little space after the last button
gr.HTML('<div style="height:60px"></div>', elem_id="bottom_spacer")
groups = {"start": g_start, "consent": g_consent, "s1": g_s1, "s2": g_s2, "done": g_done}
ALL = [state, *groups.values(), start_err, consent_err, s1_progress, *s1_texts, easiest, hardest, gap_bm,
gap_mw, gap_bm_q, gap_mw_q, overgen_q, id_box, id_saved_cb, s1_err, s2_progress, s2_source, *s2_texts, omits, contradicts, overgen, s2_err, done_md]
def show(screen: str) -> dict:
return {g: gr.update(visible=(k == screen)) for k, g in groups.items()}
def progress(st) -> str:
return f"**Document {st['pos'] + 1} of {len(st['queue'])}** · your ID: `{st['code']}`"
def enter_next(st) -> dict:
"""The current document's first screen (screen 2 directly for the validity task), or the done screen."""
if task == "validity" and st["pos"] < len(st["queue"]):
doc = C.doc_by_id(docs, st["queue"][st["pos"]])
st.update(positions=C.positions(st["code"], doc["doc_id"], systems, salt), s1=None, times={})
return enter_s2(st, doc)
return enter_s1(st)
def enter_s2(st, doc) -> dict:
view = C.screen2_view(doc, st["positions"]) # the source enters the session only now
st.update(screen="s2", view=view)
st["times"]["screen2_shown"] = C.now_iso()
return {**show("s2"), state: st, s2_progress: progress(st), s1_err: "", s2_err: "",
**{t: "" for t in s1_texts},
s2_source: _card("Source", view["source"], source=True),
**{t: _card(f"Text {L}", view["texts"][L]) for t, L in zip(s2_texts, C.LETTERS)},
omits: gr.update(value=[]), contradicts: gr.update(value=[]),
overgen: gr.update(value=[], visible=st["cohort"] == "expert"),
overgen_q: gr.update(visible=st["cohort"] == "expert")}
def save(st, s2_answers, err_box, submitted_key) -> dict:
"""Write this document's line, push it, and move on (or show the error on the current screen)."""
st["times"][submitted_key] = C.now_iso()
doc_index = next(i for i, d in enumerate(docs) if d["doc_id"] == st["view"]["doc_id"])
rec = C.build_record(st, n_docs, doc_index, s2_answers, meta)
try:
if C.append_record(out_dir, rec) and on_saved: # False = already saved from another tab
on_saved(st["code"])
except OSError as e:
st["times"].pop(submitted_key)
return {err_box: _err(f"Could not save your answers ({e}). Please try again or contact the organiser.")}
st["pos"] = C.next_index(st["queue"], C.done_ids(out_dir, st["code"]))
return enter_next(st)
def enter_s1(st) -> dict:
"""Start screen 1 for the current document, or the done screen. Clears every screen-2 field."""
out = {s2_source: "", **{t: "" for t in s2_texts}, s2_err: "", s1_err: ""}
if st["pos"] >= len(st["queue"]):
st.update(screen="done", view=None)
return {**out, **show("done"), state: st,
done_md: f"All {len(st['queue'])} documents are done (ID `{st['code']}`). Thank you! You can close "
f"this page."}
doc = C.doc_by_id(docs, st["queue"][st["pos"]])
pos = C.positions(st["code"], doc["doc_id"], systems, salt)
view = C.screen1_view(doc, pos)
st.update(screen="s1", view=view, positions=pos, s1=None, times={"screen1_shown": C.now_iso()})
return {**out, **show("s1"), state: st, s1_progress: progress(st),
**{t: _card(f"Text {L}", view["texts"][L]) for t, L in zip(s1_texts, C.LETTERS)},
easiest: gr.update(value=None), hardest: gr.update(value=None),
gap_bm: gr.update(value=None, visible=False), gap_mw: gr.update(value=None, visible=False),
gap_bm_q: gr.update(visible=False), gap_mw_q: gr.update(visible=False)}
def pull(norm):
if on_login:
try:
on_login(norm)
except Exception as e: # noqa: BLE001 -- Hub unreachable: resume from the local copy
print(f"[answer-sync] could not re-read {norm}'s answers from the repo: {e}", flush=True)
def on_code(code, key, st):
"""A returning participant: the ID must exist (signed up before, or reserved in the config)."""
if C.key_cohort(key, config) is None and C.block_key(key, config) is None:
return {start_err: _err("Wrong access key.")}
norm, err = C.validate_id(code)
if err:
return {start_err: _err(err)}
pull(norm)
if C.registration(out_dir, norm, config) is None:
return {start_err: _err(f"No participant with the ID “{norm}”. Check it, or start as new.")}
st = C.new_session(norm, config, docs, out_dir, split)
return {**show("consent"), state: st, start_err: "", consent_err: "", id_box: _id_box(norm),
id_saved_cb: gr.update(value=False)}
def on_new(code, key, st):
"""A new participant: sign the ID up (cohort from the key, next order in rotation)."""
block = None
if split and config["block_keys"]: # this Space hands out blocks by personal key only
info = C.block_key(key, config)
if info is None:
return {start_err: _err("Wrong access key. Use the personal key you were sent for this part.")}
cohort, block = info["cohort"], info["block"]
else:
cohort = C.key_cohort(key, config)
if cohort is None:
return {start_err: _err("Wrong access key.")}
norm, err = C.validate_id(code)
if err:
return {start_err: _err(err)}
taken = C.id_taken(out_dir, norm, config)
if not taken and remote_has:
try:
taken = remote_has(norm)
except Exception as e: # noqa: BLE001 -- cannot check the repo: do not risk a duplicate
print(f"[answer-sync] could not check ID {norm} in the repo: {e}", flush=True)
return {start_err: _err("Could not check this ID right now. Please try again in a minute.")}
if taken or not C.register(out_dir, norm, cohort, C.assign_order(out_dir, config, cohort), block):
return {start_err: _err(f"The ID “{norm}” is already taken. If it is yours, click “I'm coming "
f"back”; otherwise choose another.")}
if on_saved:
on_saved(norm)
st = C.new_session(norm, config, docs, out_dir, split)
return {**show("consent"), state: st, start_err: "", consent_err: "", id_box: _id_box(norm),
id_saved_cb: gr.update(value=False)}
def lost() -> dict:
"""No session: the server restarted (a Space restart) or gradio dropped an idle closed tab's state. An empty
dict would make gradio paint an 'Error' badge on every component, with no hint of what to do."""
msg = _err("This page lost its connection to the study (the page was restarted or left idle for a long "
"time). Your answers to earlier documents are saved, this one is not. Please reload this page and click "
"“I'm coming back” with your ID.")
return {consent_err: msg, s1_err: msg, s2_err: msg}
def on_consent(agreed, saved, st):
if not st:
return lost()
if st["screen"] != "consent":
return {state: st} # a stale click: nothing to do
if not saved:
return {consent_err: _err("Please save your ID first, then tick the box: without it you cannot "
"come back to where you stopped.")}
if not agreed:
return {consent_err: _err("Please tick the box to continue.")}
st["consent_at"] = C.now_iso()
return enter_next(st)
def on_rank(e, h):
"""Show the two gap questions, naming the actual letters, once easiest and hardest are set."""
hide = {gap_bm: gr.update(value=None, visible=False), gap_mw: gr.update(value=None, visible=False),
gap_bm_q: gr.update(visible=False), gap_mw_q: gr.update(visible=False)}
if e not in C.LETTERS or h not in C.LETTERS:
return {**hide, s1_err: ""}
if e == h:
return {**hide, s1_err: _err("The easiest and the hardest text must be different.")}
m = C.middle_of(e, h)
return {gap_bm: gr.update(label=C.gap_question(e, m), value=None, visible=True),
gap_mw: gr.update(label=C.gap_question(m, h), value=None, visible=True),
gap_bm_q: gr.update(value=f"**How much easier** is {e} than {m}?", visible=True),
gap_mw_q: gr.update(value=f"**How much easier** is {m} than {h}?", visible=True), s1_err: ""}
def on_s1(e, h, g1, g2, st):
if not st:
return lost()
if st["screen"] != "s1":
return {state: st}
err = C.check_screen1(e, h, g1, g2)
if err:
return {s1_err: _err(err)}
st["s1"] = C.screen1_answers(e, h, g1, g2)
if task == "simplicity":
return save(st, None, s1_err, "screen1_submitted")
st["times"]["screen1_submitted"] = C.now_iso()
return enter_s2(st, C.doc_by_id(docs, st["view"]["doc_id"]))
def on_s2(om, co, ov, st):
if not st:
return lost()
if st["screen"] != "s2":
return {state: st}
return save(st, C.screen2_answers(st["cohort"], om, co, ov), s2_err, "screen2_submitted")
code_btn.click(on_code, [code_in, key_in, state], ALL, **EVENT)
new_btn.click(on_new, [code_in, key_in, state], ALL, **EVENT)
consent_btn.click(on_consent, [consent_cb, id_saved_cb, state], ALL, **EVENT)
easiest.change(on_rank, [easiest, hardest], [gap_bm, gap_mw, gap_bm_q, gap_mw_q, s1_err], **EVENT)
hardest.change(on_rank, [easiest, hardest], [gap_bm, gap_mw, gap_bm_q, gap_mw_q, s1_err], **EVENT)
s1_btn.click(on_s1, [easiest, hardest, gap_bm, gap_mw, state], ALL, **EVENT)
s2_btn.click(on_s2, [omits, contradicts, overgen, state], ALL, **EVENT)
# On a Space, gradio's page adds a Hugging Face header (Space name, owner avatar, likes) that also loads Google
# Fonts and calls the Hub API from the participant's browser. It names the account, so drop it: space_id is only
# used for that header and for API docs snippets (off here).
demo.space_id = None
return demo
def main(argv=None):
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--data", default=os.environ.get("ANNOTATION_DATA"), help="input documents JSON")
ap.add_argument("--config", default=os.environ.get("ANNOTATION_CONFIG"),
help="default: config.yaml next to app.py, else config.example.yaml (env ANNOTATION_CONFIG)")
ap.add_argument("--out", default=os.environ.get("ANNOTATION_OUTPUT_DIR"),
help="annotation output directory (env ANNOTATION_OUTPUT_DIR; default annotations/<data file stem>)")
ap.add_argument("--layout", default=os.environ.get("ANNOTATION_LAYOUT", "horizontal"), choices=LAYOUTS,
help="texts side by side or one above another (env ANNOTATION_LAYOUT)")
ap.add_argument("--task", default=os.environ.get("ANNOTATION_TASK", "both"), choices=TASKS,
help="both screens, or only simplicity / only validity (env ANNOTATION_TASK); answers of a "
"single-task app go to a folder <data stem>-<task>")
ap.add_argument("--split", action="store_true", default=os.environ.get("ANNOTATION_SPLIT", "").strip().lower() in ("1", "true", "yes"),
help="each person gets one of the four blocks (a quarter of the documents), assigned in "
"rotation, so four people cover every document once (env ANNOTATION_SPLIT=1)")
ap.add_argument("--hub-repo", default=os.environ.get("ANNOTATION_HUB_REPO"),
help="private HF dataset repo holding config, documents and answers; --data/--config are then "
"paths inside it (env ANNOTATION_HUB_REPO; token from HF_TOKEN)")
public_env = os.environ.get("ANNOTATION_PUBLIC")
ap.add_argument("--public", action="store_true",
default=(public_env.strip().lower() not in ("", "0", "false", "no")) if public_env is not None
else bool(os.environ.get("SPACE_ID")),
help="anyone can reach the page: require a shared access key (env ANNOTATION_PUBLIC; on by "
"default on a Hugging Face Space)")
ap.add_argument("--accept-changed-data", action="store_true",
default=bool(os.environ.get("ANNOTATION_ACCEPT_CHANGED_DATA")),
help="start even if saved answers refer to another version of the documents file")
ap.add_argument("--host", default=None, help="default: gradio's (GRADIO_SERVER_NAME or 127.0.0.1)")
ap.add_argument("--port", type=int, default=None)
args = ap.parse_args(argv)
sys.stdout.reconfigure(line_buffering=True) # Space logs show each line as it happens, not at exit
if not args.data:
ap.error("--data (or ANNOTATION_DATA) is required")
stem = Path(args.data).stem + ("" if args.task == "both" else f"-{args.task}")
explicit_out = bool(args.out)
if args.hub_repo:
# always one folder per item set; without an explicit (persistent) dir, a fresh private one, so local test
# answers in the checkout are never uploaded as real ones
args.out = str(Path(args.out or tempfile.mkdtemp(prefix="annot_answers_")) / stem)
elif not args.out: # one folder per input file, so answers to different item sets never share a file
args.out = str(HERE / "annotations" / stem)
sync = None
if args.hub_repo:
import hub as H
store = H.open_store(args.hub_repo)
cache = Path(tempfile.mkdtemp(prefix="annot_private_")) # outside the app folder, never served
args.data = str(H.fetch(store, args.data, cache))
args.config = str(H.fetch(store, args.config or "config.yaml", cache))
print(f"private repo {args.hub_repo}: fetched documents and config")
elif not args.config:
args.config = str(HERE / "config.yaml") if (HERE / "config.yaml").exists() else str(HERE / "config.example.yaml")
print(f"config: {Path(args.config).name}")
docs = C.load_data(args.data)
config = C.load_config(args.config, public=args.public, access_key=os.environ.get("ANNOTATION_ACCESS_KEY"),
expert_access_key=os.environ.get("ANNOTATION_EXPERT_ACCESS_KEY"))
print(f"access key: {'required' if config['access_key'] else 'none'}; "
f"expert key: {'set' if config['expert_access_key'] else 'none (nobody joins the expert cohort)'}")
Path(args.out).mkdir(parents=True, exist_ok=True)
meta = {"data_sha256": C.file_sha256(args.data), "app_version": C.app_version()}
def refuse_stale(where: str) -> None:
stale = C.answers_from_other_data(args.out, meta["data_sha256"])
if stale and not args.accept_changed_data:
sys.exit(f"Refusing to start: saved answers {where} refer to another version of the documents file "
f"({stale} lines per code). Resume would count those documents as done. Move those answers "
f"elsewhere, or set --accept-changed-data / ANNOTATION_ACCEPT_CHANGED_DATA=1 if intended.")
refuse_stale(f"in {args.out}") # before anything local could be pushed to the repo
if args.hub_repo:
sync = H.AnswerSync(store, args.out, f"annotations/{stem}", push_local_only=explicit_out)
counts = sync.restore()
print(f"restored answers from the private repo: {counts or 'none yet'}")
def _flush_and_exit(*_): # a Space restart sends SIGTERM: push what is pending first
print("[answer-sync] shutting down, flushing pending pushes", flush=True)
sync.flush(25)
sys.exit(0)
atexit.register(sync.flush, 25)
signal.signal(signal.SIGTERM, _flush_and_exit)
if args.hub_repo:
refuse_stale(f"in the private repo under annotations/{stem}/")
print(f"{len(docs)} documents, {len(config['annotators'])} reserved IDs, writing to {Path(args.out).resolve()}")
demo = build_demo(docs, config, args.out, meta, layout=args.layout, on_saved=sync.push if sync else None,
on_login=sync.pull if sync else None, remote_has=sync.remote_has if sync else None,
task=args.task, split=args.split)
demo.queue().launch(**launch_kwargs(args.host, args.port))
if __name__ == "__main__":
main()