sawabedarain's picture
v2: LangGraph multi-agent backend + transformer knowledge tracing (ONNX)
29baa44 verified
Raw History Blame Contribute Delete
2.87 kB
"""
Bayesian Knowledge Tracing (Corbett & Anderson, 1995) with difficulty-aware guess/slip.
Each topic has four parameters:
p_init - P(L0): probability the skill is already known
p_learn - P(T): probability of learning on each practice opportunity
p_guess - P(G): probability of answering correctly while unmastered
p_slip - P(S): probability of answering incorrectly while mastered
Per-topic parameters are fitted offline (training/fit_bkt.py) and loaded from
ml/artifacts/bkt_params.json. Missing topics fall back to DEFAULT_PARAMS.
"""
import json
import os
from dataclasses import dataclass, asdict
DEFAULT_PARAMS = {'p_init': 0.10, 'p_learn': 0.15, 'p_guess': 0.22, 'p_slip': 0.10}
# Easier questions are easier to guess and harder to slip on; harder ones the opposite.
GUESS_SCALE = {'beginner': 1.3, 'intermediate': 1.0, 'advanced': 0.7}
SLIP_SCALE = {'beginner': 0.7, 'intermediate': 1.0, 'advanced': 1.4}
LEARN_SCALE = {'beginner': 0.8, 'intermediate': 1.0, 'advanced': 1.2}
@dataclass
class BKTParams:
p_init: float
p_learn: float
p_guess: float
p_slip: float
def for_difficulty(self, difficulty):
return BKTParams(
p_init=self.p_init,
p_learn=min(0.6, self.p_learn * LEARN_SCALE.get(difficulty, 1.0)),
p_guess=min(0.45, self.p_guess * GUESS_SCALE.get(difficulty, 1.0)),
p_slip=min(0.35, self.p_slip * SLIP_SCALE.get(difficulty, 1.0)),
)
class BKTModel:
def __init__(self, params_path=None):
self.params = {}
if params_path and os.path.exists(params_path):
with open(params_path) as f:
raw = json.load(f)
self.params = {slug: BKTParams(**p) for slug, p in raw.get('topics', raw).items()}
def get(self, slug):
return self.params.get(slug) or BKTParams(**DEFAULT_PARAMS)
def prior(self, slug):
return self.get(slug).p_init
def update(self, p_mastery, is_correct, slug, difficulty='intermediate'):
"""One observation -> posterior P(mastered) after the learning transition."""
p = self.get(slug).for_difficulty(difficulty)
if is_correct:
num = p_mastery * (1 - p.p_slip)
den = num + (1 - p_mastery) * p.p_guess
else:
num = p_mastery * p.p_slip
den = num + (1 - p_mastery) * (1 - p.p_guess)
posterior = num / den if den > 0 else p_mastery
return min(0.999, posterior + (1 - posterior) * p.p_learn)
def p_correct(self, p_mastery, slug, difficulty='intermediate'):
"""Predicted probability of a correct answer at the given difficulty."""
p = self.get(slug).for_difficulty(difficulty)
return p_mastery * (1 - p.p_slip) + (1 - p_mastery) * p.p_guess
def to_json(self):
return {slug: asdict(p) for slug, p in self.params.items()}