Skip to content

pykt-integration.md

Latest commit

 

History

History
79 lines (60 loc) · 2.34 KB

pykt-integration.md

File metadata and controls

79 lines (60 loc) · 2.34 KB

pyKT Integration

pyKT is a research benchmark library for deep knowledge tracing. Orchid should not try to replace that model zoo. The practical integration pattern is:

  1. Use Orchid to normalize adaptive-learning logs.
  2. Export sequences into pyKT's six-line learner format.
  3. Train or evaluate a pyKT model in the research stack.
  4. Bring predicted correctness probabilities back into Orchid for policy ranking, OPE, safety, audit, and deployment.

Export Orchid Data To pyKT

pyKT's raw sequence format stores each learner in six lines:

  1. learner id and sequence length
  2. question ids
  3. concept ids
  4. binary responses
  5. timestamps
  6. answering durations
from orchid_ranker import export_pykt_sequences

export_pykt_sequences(
    interactions,
    "data/pykt/orchid_sequences.txt",
    user_col="user_id",
    item_col="item_id",
    correct_col="correct",
    concept_col="skill_id",
    timestamp_col="timestamp",
    duration_col="elapsed_time",
)

If concepts, timestamps, or durations are unavailable, Orchid writes NA for that line's values, matching pyKT's documented raw format.

Use pyKT Predictions In Orchid

After training a pyKT model, export a prediction table with one row per user-item probability:

user_id,item_id,p_correct
u1,q1,0.82
u1,q2,0.61
u1,q3,0.74

Then use the adapter anywhere Orchid expects a KT-style predictor:

from orchid_ranker.learning_policy import KTValuePolicy
from orchid_ranker.pykt_bridge import PyKTPredictionAdapter

tracer = PyKTPredictionAdapter.from_csv("pykt_predictions.csv")
policy = KTValuePolicy(tracer, target_correct=0.70)
ranked = policy.rank("u1", ["q1", "q2", "q3"], top_k=3)

PyKTPredictionAdapter is static: it does not update a pyKT model online. That is deliberate. Use it to evaluate and serve exported model probabilities behind Orchid's policy, OPE, and safety layers. For live learning, retrain/export the pyKT model on your chosen cadence or use Orchid's native online tracers.

Round-Trip Import

from orchid_ranker.pykt_bridge import load_pykt_sequences, pykt_sequences_to_interactions

sequences = load_pykt_sequences("data/pykt/orchid_sequences.txt")
interactions = pykt_sequences_to_interactions(sequences)

This is useful for smoke tests, data audits, and verifying that exported sequence files preserve learner order and response labels.