One pipeline, eight angles, zero config — shipped as a HuggingFace Space.

The prompt — drop into mtm-resource-sweep

The deploy — 3 files, 0 config

01 · README.md (frontmatter only)

---
title: MTM Resource Sweep
emoji: 🎼
colorFrom: gold
colorTo: cyan
sdk: gradio
sdk_version: 5.13.0
app_port: 7860
suggested_hardware: l4-small
suggested_storage: small
pinned: false
license: mit
---

# MTM Resource Sweep

Weekly arXiv + monthly GitHub + quarterly foundation sweep for AMT × LLM/agent research. Zero config, scheduled inside the Space.

02 · requirements.txt (pinned)

gradio==5.13.0
huggingface_hub==0.27.0
apscheduler==3.10.4
datasets==3.2.0
basic-pitch==0.3.0
demucs==4.0.1
music21==9.3.0
verovio==4.4.0
requests==2.32.3

03 · app.py (the whole thing)

import gradio as gr
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import HfApi, CommitOperationAdd
from datetime import datetime
import json, pathlib, subprocess, sys

REPO   = "kaidjuric/mtm-resource-sweep"
SPACE  = pathlib.Path(".")
LOG    = []

def log(msg):
    line = f"[{datetime.utcnow():%H:%M:%S}] {msg}"
    LOG.append(line)
    return "\n".join(LOG[-200:])

def run_weekly_arxiv(progress=gr.Progress()):
    progress(0.1, "fetching cs.SD · cs.LG · eess.AS · cs.MM")
    out = SPACE / "reports" / f"weekly_{datetime.today():%Y-%m-%d}_arxiv.md"
    out.parent.mkdir(exist_ok=True)
    # worker lives in mtm_sweep/arxiv_sweep.py — called as a module
    subprocess.run([sys.executable, "-m", "mtm_sweep.arxiv_sweep",
                    "--out", str(out)], check=True)
    progress(1.0, "pushing to Space")
    HfApi().create_commit(REPO, [CommitOperationAdd("reports/"+out.name,
                       pathlib.Path(out).read_bytes())],
                       commit_message=f"sweep: {out.name}")
    return log(f"weekly sweep → {out.name}")

with gr.Blocks(title="MTM Resource Sweep", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🎼 MTM Resource Sweep\nZero-config. Weekly arXiv · monthly GitHub · quarterly foundation.")
    run  = gr.Button("Run weekly sweep now", variant="primary")
    logb = gr.Textbox(label="live cycle log", lines=20, interactive=False)
    run.click(run_weekly_arxiv, outputs=logb, every=2)

# schedule the worker — Mon 09:00 CET = 08:00 UTC (winter) / 07:00 UTC (DST)
sched = BackgroundScheduler(daemon=True)
sched.add_job(run_weekly_arxiv, "cron", day_of_week="mon", hour=7, minute=0)
sched.start()

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)

Tool stack Gradio 5.13 + APScheduler

Benchmarking segment

Verdict

The prompt — drop into mtm-resource-sweep

The UI — 4 tabs, 1 screen

Tool stack Gradio 5.13 · Soft theme

Benchmarking segment

Verdict

The prompt — drop into mtm-resource-sweep

The pipeline — model stack

Tool stack all on HF Hub

Benchmarking segment

Verdict

The prompt — drop into mtm-resource-sweep

The pipeline — worker dispatch

Tool stack asyncio + STATUS.json

Benchmarking segment

Verdict

The prompt — drop into mtm-resource-sweep

The pipeline — scheduler state machine

Tool stack L4 never sleeps

Benchmarking segment

Verdict

The prompt — drop into mtm-resource-sweep

The library — 8 prompts, 1 module

mtm_sweep/prompts.py

from string import Template
from pathlib import Path
import hashlib, json

# === ARXIV ================================================================
ARXIV_SWEEP_COMPACT = """Sweep arXiv cs.SD, cs.LG, eess.AS, cs.MM from $FROM to $TO.
For each new paper touching AMT, source separation, OMR, or score rendering:
  - arxiv_id · title · authors
  - 1-line relevance tag
Return Markdown with ≥5 entries, each with a working arXiv link."""

ARXIV_SWEEP_STANDARD = """Sweep arXiv cs.SD, cs.LG, eess.AS, cs.MM from $FROM to $TO for papers touching
**Automatic Music Transcription (AMT)**, source separation, OMR, or score
rendering. For each entry:
  1. arXiv ID + title + authors
  2. Abstract snippet (≤40 words)
  3. 1-line relevance tag
  4. Direct https://arxiv.org/abs/... link
  5. Code repo link (if any) — GitHub, HF Space, or paper page
Sort by recency. Cap at 15. Cross-check `data/sweep_history.parquet` for dups.
Output: weekly_$DATE_arxiv.md. No claims without a primary URL."""

ARXIV_SWEEP_DEEP = """Deep sweep arXiv cs.SD, cs.LG, eess.AS, cs.MM from $FROM to $TO.
For each paper touching AMT/OMR/separation/score rendering:
  1. arXiv ID · title · authors · date
  2. Abstract (verbatim, ≤40 words)
  3. Method (1 paragraph)
  4. Dataset + eval
  5. License / weight availability
  6. Code repo (GitHub / HF Space) + last commit
  7. Reproducibility (≤24h on 1 GPU? docker? HF Space demo?)
  8. Relevance score 0-5 against the mtm pipeline
  9. Hot take: should we drop, watch, integrate, or replace a stage?
Sort by relevance. Cap at 20. Output: weekly_$DATE_arxiv.md + sweep.json."""

# === GITHUB ===============================================================
GITHUB_PULSE_STANDARD = """For each repo in $REPOS list:
  - latest release tag + date + breaking changes vs prior
  - stars, forks, last_commit, last_release
  - 1-line "is this still on the stack?" verdict
If a repo has a new release > 7 days old, flag it as a watch item.
Output: github_$YYYY_MM.md. No claims without a release-page URL."""

# === FOUNDATION ============================================================
FOUNDATION_AUDIT_STANDARD = """Quarterly review of trusted_packages.txt.
For each package:
  1. Latest version + release date
  2. CVE check (NVD + GitHub Advisory) — any open > 30 days?
  3. License change since last quarter?
  4. Still imported? (grep -r in setup/ + transcriber-app/)
  5. Maintenance health (commits/month, open issues, last response)
Then run: python3 -c "import ast; ast.parse(open('setup/pipeline_starter.py').read())"
Output: foundation_$YYYY_Qn.md. Cite every change with a primary source."""

# === RANKER LLM ===========================================================
RANKER_HOT_TAKE = """Given these $N papers:
$PAPERS

Write 1 hot-take paragraph per paper. Constraint: ≤40 words, no hedging,
no "could potentially". State the position. Score 0-5 for fit into the
existing mtm pipeline (basic-pitch · demucs · music21 · verovio · MT3).
Return JSON: {papers: [{id, hot_take, score}]}.
Temperature: 0.2. Top-p: 0.9. Max tokens: 800."""

# === DIGEST EMAIL =========================================================
DIGEST_EMAIL = """Subject: mtm weekly · $DATE · $N papers

Top 3 by relevance:
$TOP3

Watch items:
$WATCH

Reproduce locally: python3 -m mtm_sweep.arxiv_sweep --out reports/
Full report: $URL
Reply STOP to unsubscribe."""

# === FAILURE HANDLER ======================================================
FAILURE_HANDLER = """Worker $WORKER failed twice. Last error:
$ERROR

Diagnose: transient (network, 429, 5xx) or systematic (schema drift, bad URL)?
If transient: back off 1h, retry. If systematic: write STATUS.json with
$WORKER=fail and email the owner. Do not retry more than 3 times in 24h."""

def render(name: str, **kwargs) -> str:
    tmpl = globals()[name]
    return Template(tmpl).substitute(**kwargs)

def hash_prompt(name: str) -> str:
    return hashlib.sha256(globals()[name].encode()).hexdigest()[:12]

Tool stack string.Template · hash for diff

Benchmarking segment

Verdict

The prompt — drop into mtm-resource-sweep

The bench — every number, end-to-end

Tool stack verdict

Composite benchmark

Verdict