bihungba1101/slam-es-en-knowledge-tracing

SLAM Spanish-English Knowledge Tracing

This repository contains a Hugging Face-compatible conversion of the Spanish-English track from the 2018 Second Language Acquisition Modeling (SLAM) shared task.

Each row represents one learner exercise attempt. The original token-level mistake labels are retained in token_labels: 0 means OK and 1 means a mistake. The correct field is strict whole-attempt correctness: it is 1 only when every token label is 0.

This dataset is suitable for longitudinal knowledge tracing because it records repeated interactions for anonymous learners, relative elapsed time, response correctness, challenge type, session type, and response time. Challenge types are not restricted to multiple-choice questions; they include language-learning activities such as listening, reverse tapping, and reverse translation.

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 Spanish-English data. For the prompt Good morning!, the same stored reference answer Buenos días occurs once with labels [0, 0] and elsewhere with labels [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.

Dataset schema

The following example shows one converted interaction:

{
  "user_id": "+H9QWAV4",
  "interaction_index": 4,
  "exercise_id": "Ii4fn+U101",
  "prompt": "The knife is black.",
  "country": "CA",
  "client": "ios",
  "session": "lesson",
  "format": "reverse_translate",
  "days": 13.424,
  "response_time_seconds": 25.0,
  "token_ids": [
    "Ii4fn+U10101",
    "Ii4fn+U10102",
    "Ii4fn+U10103",
    "Ii4fn+U10104"
  ],
  "reference_tokens": ["El", "cuchillo", "es", "negro"],
  "pos_tags": ["DET", "NOUN", "VERB", "ADJ"],
  "morphological_features": [
    "Definite=Def|Gender=Masc|Number=Sing|PronType=Art|fPOS=DET++",
    "Gender=Masc|Number=Sing|fPOS=NOUN++",
    "Mood=Ind|Number=Sing|Person=3|Tense=Pres|VerbForm=Fin|fPOS=VERB++",
    "Gender=Masc|Number=Sing|fPOS=ADJ++"
  ],
  "dependency_relations": ["det", "nsubj", "cop", "ROOT"],
  "dependency_heads": [2, 4, 4, 0],
  "token_labels": [1, 1, 0, 1],
  "token_accuracy": 0.25,
  "correct": 0,
  "prompt_available": True
}

Each token-level list is positionally aligned. For example, reference_tokens[2] is "es", its corresponding token_ids[2] is "Ii4fn+U10103", its pos_tags[2] value is "VERB", its dependency_relations[2] value is "cop", and its token_labels[2] value is 0 (OK).

The listed tokens and linguistic annotations describe the closest correct reference answer selected by Duolingo's alignment system, not the learner's raw response, which is not provided. For example, DET means determiner, NOUN means noun, and ROOT identifies the syntactic root. A dependency head is a one-based token position; 0 means that the token is the root.

Feature reference

Feature	Type	Description	Example
user_id	string	Anonymous learner identifier. Rows with the same value belong to the same learner.	"+H9QWAV4"
interaction_index	integer	Sequential attempt number for that learner within the source split.	4
exercise_id	string	Identifier derived from the reference token IDs for this attempt. It identifies an attempt instance and should not automatically be treated as a stable question/content ID.	"Ii4fn+U101"
prompt	string or null	English source prompt when the original interaction provides one. It is often absent for listening activities.	"The knife is black."
country	string	Country metadata associated with the learner.	"CA"
client	string	Application or device platform used for the interaction.	"ios"
session	string	Learning session type, such as lesson or practice.	"lesson"
format	string	Challenge type, such as listen, reverse_tap, or reverse_translate.	"reverse_translate"
days	float	Learner-relative elapsed time in days from the beginning of the learner's observed activity timeline. It is not a calendar date.	13.424
response_time_seconds	float or null	Time spent on the attempt in seconds. Missing source values are null.	25.0
token_ids	list of strings	Source identifiers for tokens in the closest correct reference answer.	["Ii4fn+U10101", "Ii4fn+U10102"]
reference_tokens	list of strings	Tokens in the closest correct reference answer; these are not the learner's raw response.	["El", "cuchillo", "es", "negro"]
pos_tags	list of strings	Universal part-of-speech tags aligned with reference_tokens.	["DET", "NOUN", "VERB", "ADJ"]
morphological_features	list of strings	Morphological annotations aligned with reference_tokens, including features such as gender, number, tense, person, and fine-grained POS.	`["Gender=Masc
dependency_relations	list of strings	Syntactic dependency relation for each reference token, such as det, nsubj, cop, or ROOT.	["det", "nsubj", "cop", "ROOT"]
dependency_heads	list of integers	One-based syntactic head position for each reference token; 0 identifies the root.	[2, 4, 4, 0]
token_labels	list of integers	Official mistake labels aligned with reference_tokens; 0 means OK and 1 means a mistake.	[1, 1, 0, 1]
token_accuracy	float	Fraction of reference positions without a learner mistake: 1 - mean(token_labels).	0.25
correct	integer	Strict whole-attempt target; 1 only when every token_labels value is 0, otherwise 0.	0
prompt_available	boolean	Whether a prompt is available for this interaction.	True

Understanding time and learner histories

The days feature is learner-relative elapsed time. It indicates how many days had passed since the beginning of that learner's recorded activity timeline. It is not an account creation date, a calendar date, or a global timestamp shared by all learners.

For example, the following values indicate activity near the beginning of a learner timeline, followed by activity roughly 13 days later:

days:0.009
days:13.424
days:14.710

The time gap between consecutive interactions is calculated as:

time_gap_days = days_current - days_previous

For example, 14.710 - 13.424 = 1.286 days, or approximately 30.86 hours. This supports modelling practice spacing, forgetting effects, and learning progression.

The first interaction visible in a particular split does not necessarily have days = 0. A learner may have earlier interactions in another source split, so a validation or test record can begin at a later relative time.

Correctness and modelling targets

The source data provides mistake labels at token positions in the closest correct reference answer. The official semantics are 0 = OK and 1 = mistake. The conversion preserves these labels and derives two summary targets:

• token_accuracy: the fraction of aligned reference positions without a learner mistake.
• correct: strict interaction-level correctness.

The strict interaction label is defined as:

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

For example:

token_labels	token_accuracy	correct
[0, 0, 0]	1.00	1
[1, 1, 0, 1]	0.25	0
[1, 1]	0.00	0

This supports two related tasks:

1. Interaction-level knowledge tracing: predict whether the learner will complete the next exercise correctly.
2. Token-level knowledge tracing: predict whether the learner will make a mistake at each reference-token position.

A typical interaction-level KT objective is:

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

The most important fields for this task are user_id, days, interaction_index, format, prompt, correct, response_time_seconds, and session.

Example learner sequence

Rows can be grouped by user_id and ordered by days and interaction_index:

interaction_index	days	format	prompt	correct
1	0.009	listen	null	0
2	0.009	listen	null	1
3	0.009	reverse_translate	The bread	1
4	13.424	reverse_translate	The knife is black.	0
5	13.424	reverse_tap	I have plates.	1

A knowledge-tracing model can use interactions 1 through 4 to predict the outcome of interaction 5.

Linguistic annotation example

The original files contain rows such as:

Lmzbsr1h0201  El       DET   Definite=Def|Gender=Masc|Number=Sing|PronType=Art|fPOS=DET++  det   2
Lmzbsr1h0202  espejo   NOUN  Gender=Masc|Number=Sing|fPOS=NOUN++                       ROOT  0

In the converted dataset, these rows are stored inside one interaction as aligned lists:

{
  "exercise_id": "Lmzbsr1h02",
  "reference_tokens": ["El", "espejo"],
  "pos_tags": ["DET", "NOUN"],
  "morphological_features": [
    "Definite=Def|Gender=Masc|Number=Sing|PronType=Art|fPOS=DET++",
    "Gender=Masc|Number=Sing|fPOS=NOUN++"
  ],
  "dependency_relations": ["det", "ROOT"],
  "dependency_heads": [2, 0],
  "token_labels": [0, 0]
}

Here, El is a masculine singular definite determiner attached to token position 2, while espejo is a masculine singular noun and the syntactic root. The correctness labels shown above are illustrative of the converted row; use the published row values for the exact labels in a particular split.

Strict correctness balance and time analysis

Using the official semantics (0 = OK, 1 = mistake), the strict target is moderately imbalanced toward correct attempts:

Split	Strict correct	Incorrect (one or more mistakes)	Correct rate
train	532,591	199,305	72.77%
validation	64,191	31,812	66.86%
test	61,462	31,683	65.99%
All splits	658,244	262,800	71.47%

Across learner-relative time, the unadjusted strict-correct rate declines:

Learner-relative day	Attempts	Correct rate	Average tokens
<= 1	180,339	77.54%	2.39
1-3	158,933	73.03%	2.68
3-7	207,698	70.53%	2.80
7-14	215,704	69.02%	2.90
> 14	158,370	67.54%	3.03

This decline must not be interpreted directly as learners becoming worse. Later interactions contain longer responses and a larger share of difficult reverse_translate exercises. Strict correctness is inherently length-sensitive because an attempt is incorrect when any token position has a mistake. Correct rates also differ substantially by format: reverse_tap 88.10%, listen 72.03%, and reverse_translate 52.50%. Models and evaluations should therefore control for response length, format, session, and learner-specific history.

Splits

• train: source training split.
• validation: source development split.
• test: source test split.

The converted split sizes are:

Split	Interactions
train	731,896
validation	96,003
test	93,145

The source files contain approximately 2,640 anonymous learners per split. The same learner IDs occur across the supplied splits, so these splits support longitudinal evaluation but are not learner-disjoint cold-start splits. Create new user-disjoint splits when evaluating generalization to unseen learners.

Load

from datasets import load_dataset

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

Important: the supplied source splits reuse learner IDs across splits. They are not learner-disjoint cold-start splits.

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 in this repository. The converter 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 all linguistic annotations, and derives strict interaction-level correctness.