title	PLM Factory — Perceptual Learning Drills
emoji	🧠
colorFrom	blue
colorTo	purple
sdk	gradio
sdk_version	6.20.0
app_file	app.py
pinned	false
license	mit
short_description	Adaptive perceptual learning drills (MATH 223 demo)

🧠 PLM Factory

A multi-agent system (built on smolagents) that generates Perceptual Learning Modules — rapid, timed, high-repetition classification drills that build fast visual/pattern intuition — for three University of Arizona courses:

Course	Perceptual skill trained	Item engine (ground-truth oracle)
MATH 223 — Vector Calculus	Read a vector field plot, instantly judge div/curl sign	sympy (symbolic — exact keys)
CHEM 241A — Organic Chemistry	Recognize functional groups, stereo, mechanisms at a glance	rdkit (SMARTS / CIP)
ARH 201/202 — Western Art surveys	Classify style/period/artist from a fragment	WikiArt labels + museum APIs

This Space hosts two tabs:

1. MATH223 Student Demo — the FieldSense drill: timed 3-way classification of divergence/curl sign with ARTS adaptive sequencing (accuracy and response time; categories retire at 4 consecutive fast-and-correct answers), per Kellman, Massey & Son (2010), Topics in Cognitive Science.
2. Agent Debug (dev) — pick a course + concept and watch a real smolagents CodeAgent (Qwen2.5-Coder-32B-Instruct) generate a live item end to end, with its code trace shown so you can verify it calls the course's sanctioned generator tool rather than asserting an answer itself. Deliberately built as fixed dropdowns rather than smolagents' built-in free-text GradioUI chat, since an open-ended chat into a code-executing agent is a real prompt-injection surface on a public Space.

Architecture (v2)

Three agents: Retrieval Agent (Chroma ×3 collections, BAAI/bge-m3 embeddings via the HF Inference API) + Research Agent (DuckDuckGo web search → cached, per-topic concept briefs; Qwen2.5-Coder-32B-Instruct) + a single PLM Agent (CodeAgent, course as a parameter; Qwen2.5-Coder-32B-Instruct) whose chain-of-thought only picks category / difficulty / instance parameters — all item content and answer keys come from deterministic code oracles (sympy / RDKit / WikiArt's own curated labels), never LLM assertion, enforced live via a step_callbacks guard. Full design: ARCHITECTURE.md.

Repo layout

app.py                        # this Space: student demo tab + agent debug tab
src/config.py                 # model assignments, Chroma paths, ARTS constants
src/plm_core/arts.py          # ARTS adaptive sequencing tracker (deterministic)
src/plm_core/plm_agent.py     # PLM Agent (CodeAgent, course as a parameter)
src/plm_core/retrieval.py     # Retrieval Agent + retrieve_course_context tool
src/plm_core/research.py      # Research Agent (web -> cached per-topic briefs)
src/plm_core/embeddings.py    # bge-m3 embeddings via HF Inference API
src/plm_core/ingest.py        # one-time textbook PDF -> Chroma ingestion (local-only)
src/plm_core/wikiart_cache.py # one-time WikiArt exemplar caching (local-only)
src/generators/math223.py     # sympy-oracle FieldSense item generator
src/generators/chem241a.py    # RDKit-oracle GroupSense/StereoSense item generator
src/generators/arh.py         # WikiArt-label-oracle StyleSense item generator
data/chroma/                  # pre-built Chroma DB (math223, chem241, arh collections)
data/wikiart_cache/           # cached WikiArt exemplar images + manifest
ARCHITECTURE.md               # confirmed v2 system design
PROPOSAL.md                   # superseded v1 (kept for item schema / ARTS detail)
requirements.txt              # what's actually installed on this Space
requirements-full.txt         # full project stack incl. local-only ingestion/observability