bihungba1101/slam-en-es-knowledge-tracing

SLAM English-Spanish Knowledge Tracing

This repository contains a Hugging Face-compatible conversion of the English-Spanish track from the 2018 Second Language Acquisition Modeling (SLAM) shared task. Each row represents one learner exercise attempt. Source reference-token, linguistic, and official mistake-label annotations are preserved.

The dataset supports longitudinal knowledge tracing: the same anonymous learner can have many interactions over multiple relative days, with challenge type, session type, response time, prompt text when available, and token-level outcomes.

Prompt, reference answer, and learner response

These are three distinct concepts:

• prompt is the text shown to the learner when text is available. Listening exercises commonly have no text prompt.
• reference_tokens contains the tokens of the acceptable correct answer that Duolingo's finite-state alignment system selected as most similar to the learner's submission.
• The learner's literal submitted text is not included in the SLAM release and cannot be reconstructed from reference_tokens and token_labels.

The distinction is visible in the raw English-Spanish data. The same learner received the prompt Yo soy un niño. twice. Both records store the same reference answer, I am a boy; the first has labels [0, 0, 0, 0], while the second has [1, 1, 1, 1]. If the stored words were the literal learner response, the identical words could not be both fully matched and fully mistaken. Instead, the words identify the aligned correct answer, while the labels describe mismatches in an unavailable learner submission.

This interpretation is also stated explicitly in the official SLAM 2018 task specification: the release provides token-level labels and dependency parses for the most similar correct answer to each submission and does not provide actual student responses. Therefore, reference_tokens means aligned correct-reference tokens, not response tokens.

Splits

Split	Interactions	Learners	Prompt coverage
train	824,012	2,593	72.2%
validation	115,770	2,592	70.7%
test	114,586	2,593	70.2%

The supplied source splits reuse learner IDs across splits. They are not learner-disjoint cold-start splits. Create new user-disjoint splits when evaluating generalization to unseen learners.

Schema

Each token-level list is positionally aligned. For example:

{
    "user_id": "XEinXf5+",
    "interaction_index": 2,
    "exercise_id": "rsAkJBG002",
    "prompt": "Hola, buenas tardes.",
    "country": "CO",
    "client": "web",
    "session": "lesson",
    "format": "reverse_translate",
    "days": 13.693,
    "response_time_seconds": 15.0,
    "token_ids": [
        "rsAkJBG00201",
        "rsAkJBG00202",
        "rsAkJBG00203"
    ],
    "reference_tokens": ["Hello", "good", "afternoon"],
    "pos_tags": ["INTJ", "ADJ", "NOUN"],
    "morphological_features": [
        "fPOS=INTJ++UH",
        "Degree=Pos|fPOS=ADJ++JJ",
        "Number=Sing|fPOS=NOUN++NN"
    ],
    "dependency_relations": ["ROOT", "amod", "nmod:tmod"],
    "dependency_heads": [0, 4, 1],
    "token_labels": [0, 0, 0],
    "token_accuracy": 1.0,
    "correct": 1,
    "prompt_available": True
}

Feature	Meaning	Example
user_id	Anonymous learner identifier used to group interactions.	"XEinXf5+"
interaction_index	Sequential attempt number for the learner within the source split.	2
exercise_id	Identifier derived from the reference token IDs for this attempt.	"rsAkJBG002"
prompt	Source prompt when available; often absent for listening activities.	"Hola, buenas tardes."
country	Learner country metadata.	"CO"
client	Application or device platform.	"web"
session	Session type, such as lesson or practice.	"lesson"
format	Challenge type, such as listen, reverse_tap, or reverse_translate.	"reverse_translate"
days	Learner-relative elapsed time in days, not a calendar date.	13.693
response_time_seconds	Time spent on the attempt; may be null.	15.0
token_ids	Source IDs for tokens in the closest correct reference answer.	["rsAkJBG00201", ...]
reference_tokens	Tokens in the closest correct reference answer; these are not the learner's raw response.	["Hello", "good", "afternoon"]
pos_tags	Universal POS tags aligned with the reference tokens.	["INTJ", "ADJ", "NOUN"]
morphological_features	Morphological and fine-grained POS annotations.	`["Degree=Pos
dependency_relations	Syntactic relations such as ROOT, amod, and nmod:tmod.	["ROOT", "amod", "nmod:tmod"]
dependency_heads	One-based syntactic head positions; 0 identifies the root.	[0, 4, 1]
token_labels	Official mistake labels aligned with reference_tokens: 0 means OK and 1 means a mistake.	[0, 0, 0]
token_accuracy	Fraction of reference positions without a learner mistake: 1 - mean(token_labels).	1.0
correct	Strict whole-attempt correctness; 1 only if every token label is 0.	1
prompt_available	Whether prompt text is available.	True

The days field is the number of days since the learner started learning Spanish on Duolingo. Time gaps can be calculated as days_current - days_previous. A first row in a particular split may have a nonzero value because the learner may have earlier records in another split. Null response times are preserved, and negative times caused by known client logging errors are converted to null.

The listed tokens and linguistic annotations describe the closest correct reference answer selected by Duolingo's alignment system. The learner's raw response is not provided. token_labels indicates where that response was missing or misspelled relative to the aligned reference.

Knowledge-tracing targets

The conversion supports interaction-level KT using correct and token-level mistake prediction using token_labels. Strict correctness is derived as:

token_accuracy = 1 - mean(token_labels)
correct = 1 if every token label is 0; otherwise correct = 0

A typical interaction-level objective is:

P(correct[t+1] = 1 | learner history through interaction t, target exercise[t+1])

Load

from datasets import load_dataset

dataset = load_dataset("bihungba1101/slam-en-es-knowledge-tracing")

Citation and license

See CITING.md and LICENSE.md. This conversion does not change the source dataset license or citation requirements.

Conversion

The Parquet files were generated by convert_to_hf.py. It groups source reference-token rows into one interaction row, joins development and test labels from their .key files, preserves official token-level mistake labels and linguistic annotations, and derives strict interaction-level correctness.