{cfg['title']}{active_badge}
{cfg['subtitle']}
{cfg['blurb']}
{cfg['delta']}
{cfg['deltatxt']}
{cfg['stats']}
"""
if not is_active:
card_html += f'
Open trainer ↗'
card_html += "
"
st.markdown(card_html, unsafe_allow_html=True)
st.markdown("")
st.markdown("---")
# ---------------------------------------------------------------------------
# Active-task training panel
# ---------------------------------------------------------------------------
active_cfg = TASKS[LOCAL_TASK]
# ---------------------------------------------------------------------------
# Pre-staged sample-run artifacts (additive, only show if files exist).
# Lets judges see real output without having to click Run.
# ---------------------------------------------------------------------------
have_sample = (
os.path.exists(SAMPLE_LOG_FILE)
or os.path.exists(SAMPLE_BEFORE_AFTER_PNG)
or os.path.exists(SAMPLE_RESULTS_JSON)
)
if have_sample:
st.subheader(f"📊 {active_cfg['title']} · sample run (already trained)")
st.markdown(
f"Below is the recorded output of a successful **{LOCAL_TASK}** "
f"training run on this Space — the same run that produced the "
f"`{active_cfg['delta']}` headline number. You can read the log + "
f"plots without having to launch a new run."
)
if os.path.exists(SAMPLE_RESULTS_JSON):
try:
r = json.load(open(SAMPLE_RESULTS_JSON))
c1, c2, c3 = st.columns(3)
c1.metric("Baseline mean", f"{r.get('baseline_mean', 0):.3f}")
c2.metric("Trained mean", f"{r.get('trained_mean', 0):.3f}")
c3.metric("Improvement", f"{r.get('improvement', 0):+.3f}")
except Exception:
pass
pcols = st.columns(2)
if os.path.exists(SAMPLE_BEFORE_AFTER_PNG):
pcols[0].image(SAMPLE_BEFORE_AFTER_PNG, caption="Before vs After SFT (sample run)")
if os.path.exists(SAMPLE_REWARD_PNG):
pcols[1].image(SAMPLE_REWARD_PNG, caption="Training Loss (sample run)")
if os.path.exists(SAMPLE_LOG_FILE):
with st.expander("📜 Sample training log (full output)"):
try:
st.text(open(SAMPLE_LOG_FILE).read())
except Exception as e:
st.error(f"Could not read sample log: {e}")
st.markdown("---")
st.subheader(f"▶ {active_cfg['title']} · run training here")
st.markdown(
f"Trains via SFT → GRPO on the live `{LOCAL_TASK}_review` task of the "
f"SecureReview environment. Reward comes from the live grader — no static dataset."
)
done = os.path.exists(DONE_FILE)
log_present = os.path.exists(LOG_FILE)
training_alive = is_training_alive()
ongoing = log_present and not done
if not ongoing and not done:
if st.button("▶ Run Training", type="primary"):
for p in (DONE_FILE, LOG_FILE, PID_FILE):
if os.path.exists(p):
try: os.remove(p)
except OSError: pass
threading.Thread(target=_run, daemon=True).start()
time.sleep(1)
st.rerun()
elif ongoing:
if training_alive:
st.info("Training in progress. Page auto-refreshes every 10 s — safe to close your laptop.")
else:
st.warning(
"Training process not detected (container may have restarted). "
"If the log below has stopped growing, click **Restart Training**."
)
log_content = ""
if log_present:
with open(LOG_FILE) as f:
log_content = f.read()
st.text_area("Training Log", log_content, height=460)
col_a, col_b = st.columns(2)
if not training_alive:
if col_a.button("Restart Training (clears progress)"):
for p in (DONE_FILE, LOG_FILE, PID_FILE):
if os.path.exists(p):
try: os.remove(p)
except OSError: pass
threading.Thread(target=_run, daemon=True).start()
time.sleep(1)
st.rerun()
if training_alive:
time.sleep(10)
st.rerun()
else:
st.success("Training complete!")
if os.path.exists(RESULTS_FILE):
with open(RESULTS_FILE) as f:
r = json.load(f)
c1, c2, c3 = st.columns(3)
c1.metric("Baseline mean", f"{r['baseline_mean']:.3f}")
c2.metric("Trained mean", f"{r['trained_mean']:.3f}")
c3.metric("Improvement", f"{r['improvement']:+.3f}")
col1, col2 = st.columns(2)
if os.path.exists(f"{PLOTS_DIR}/reward_curve.png"):
col1.image(f"{PLOTS_DIR}/reward_curve.png", caption="Reward Curve")
if os.path.exists(f"{PLOTS_DIR}/before_after.png"):
col2.image(f"{PLOTS_DIR}/before_after.png", caption="Before vs After")
if log_present:
with st.expander("Training log"):
with open(LOG_FILE) as f:
st.text(f.read())
if st.button("Reset & Run Again"):
for p in (DONE_FILE, LOG_FILE, PID_FILE):
if os.path.exists(p):
try: os.remove(p)
except OSError: pass
st.rerun()