TRACE-BN: Transferring Bangla-English Tutoring Behavior to a Sub-1B Offline Language Model

TRACE-BN is a structured bilingual tutoring dataset of 4,099 teacher traces grounding Bangla-English grammar instruction. Each trace is a 7-field JSON object capturing the full pedagogical arc from word-level gloss alignment through to a targeted practice exercise.

---

Overview

TRACE-BN addresses the low-resource challenge of building offline, on-device Bangla-to-English tutoring for secondary school learners (NCTB Classes 9-10) in Bangladesh. Synthetically generated by Gemini 3.5 Flash-Lite, filtered through a rigorous 3-tier automated pipeline, and validated by three bilingual language educators.

Property	Value
Total traces (after filtering)	4,099
Train / Val / Test	3,667 / 367 / 432
Source language	Bangla
Target language	English
CEFR levels	A1, A2
Fields per trace	7
Curriculum scope	13 grammar units + 15 topic clusters
Human validation avg.	4.71 / 5.0 (3 educators, 100 traces)

---

Data Fields

Field	Type	Description
id	string	UUID-format unique record identifier
topic	string	Thematic topic cluster
cefr_level	string	Target CEFR level: A1 or A2
bangla_sentence	string	Source Bangla input sentence
word_gloss	list[dict]	Word-by-word gloss (bn, en, pos)
literal_translation	string	Word-for-word English reflecting Bangla syntax
natural_translation	string	Fluent idiomatic English target
grammar_notes	list[str]	Contrastive grammar notes in Bangla
common_mistake	string	Typical L1-interference error
practice_question	string	Follow-up practice prompt in Bangla
practice_answer	string	Correct English answer
split	string	train / validation / test

---

Data Splits

Split	Count	Purpose
train	3,667	LoRA fine-tuning supervision
validation	367	Hyperparameter selection
test	432	Held-out evaluation (weather + idiom clusters)
Full corpus	4,099	data/trace_bn_full_4099.jsonl

---

Construction and Quality

Synthesis Pipeline

1. Source sentences from NCTB Class 9-10 English Grammar textbook via OCR
2. Traces synthesized by Gemini 3.5 Flash-Lite with structured few-shot prompts
3. 3-tier automated filtering (351/4,450 removed, 7.9%):

Tier	Criterion	Removed
Structural Validation	JSON parse, 7 required fields, gloss token ratio	180
Script Integrity	Bangla Unicode U+0980-U+09FF verified	45
Semantic Deduplication	MiniLM cosine similarity > 0.92	126

Human Validation (3 educators, 100 traces, 1-5 Likert)

Dimension	Mean +/- SD
Translation Quality	4.81 +/- 0.58
Grammar Explanation	4.81 +/- 0.62
Learner-Mistake Plausibility	4.79 +/- 0.68
Practice-Question Alignment	4.44 +/- 1.17
Overall Pedagogical Rating	4.71 +/- 0.59

95.7% of audited traces rated >= 4. Zero factual or grammatical errors found.

---

Benchmark Results

Fine-tuned Qwen3-0.6B with 4-bit LoRA (r=16, alpha=32) on NVIDIA T4 GPU, 3 epochs (621 steps):

Model	Params	Schema Validity	chrF++	BLEU	Throughput (tok/s)
Gemma-4 E2B	2.3B eff.	90.5%	50.91	33.54	45.97
TigerLLM-1B	1.0B	67.1%	22.23	8.21	48.79
Llama-3.2-1B-Instruct	1.0B	12.7%	13.21	1.95	73.11
Qwen3-0.6B (Base)	0.6B	85.4%	15.28	4.52	90.89
Qwen3-0.6B (TRACE-BN LoRA)	0.6B	95.8%	34.77	21.03	75.47

+19.49 chrF++ improvement (p < 0.05, paired bootstrap, 1,000 iterations)

Dual-Judge Pedagogical Evaluation (0-4 scale, GPT-4o-mini + DeepSeek-V4-Flash-0731)

Dimension	Base	Tuned	Delta	kappa
Word Gloss Accuracy	0.10	1.48	+1.38	0.788
Literal Translation	0.21	1.33	+1.12	0.754
Natural Translation	0.26	1.31	+1.05	0.803
Grammar Explanation	0.27	1.48	+1.21	0.489
Learner-Error Diagnosis	0.23	1.47	+1.24	0.585
Practice Alignment	0.16	1.17	+1.01	0.529
Overall Tutoring Quality	0.13	1.14	+1.01	0.700

---

Quickstart

from datasets import load_dataset

ds = load_dataset("RaiyanKhaan/Trace-BN")
print(ds["train"][0]["bangla_sentence"])
print(ds["train"][0]["natural_translation"])
print(ds["train"][0]["grammar_notes"])

---

Repository Layout

RaiyanKhaan/Trace-BN/
data/                       <- HF-native JSONL splits
  train.jsonl               <- 3,667 training traces
  validation.jsonl          <- 367 validation traces
  test.jsonl                <- 432 held-out test traces
  trace_bn_full_4099.jsonl  <- Complete corpus
  ablations/                <- Multi-task ablation subsets
  schema/
    trace_schema.json       <- Formal JSON Schema (Draft 2020-12)
experiments/
  predictions/              <- Raw outputs for 5 model baselines
  evaluations/              <- Benchmark, dual-judge, ablation metrics
  human_evaluation/         <- Educator audit sheets + agreement
scripts/
  data_generation/          <- Prompt templates + 3-tier filter
  training/                 <- LoRA fine-tuning + GGUF export
  evaluation/               <- chrF++, BLEU, dual-judge, bootstrap
notebooks/
  TRACE_BN_Qwen3_LoRA_FineTuning.ipynb
figures/                    <- Paper diagrams and performance charts
TRACE_BN_Paper.pdf
CITATION.bib
LICENSE

---

Citation

@inproceedings{reza2026tracebn,
  title={{TRACE-BN}: Transferring Bangla-English Tutoring Behavior to a Sub-1B Offline Language Model},
  author={Reza, Khan Raiyan Ibne and Maria, Sanjana Aktar and Abdullah, Mohammad Tushar and Leen, Asfee Bhuiyan and Nimi, Sumaiya Tabassum},
  booktitle={Proceedings of the IEEE International Conference on Computer Science and Data Engineering (CSDE)},
  year={2026}
}

---

License and Contact

Released under the MIT License.

Authors - Department of Electrical and Computer Engineering, North South University, Dhaka, Bangladesh:

• Khan Raiyan Ibne Reza (raiyan.reza@northsouth.edu)
• Sanjana Aktar Maria (sanjana.maria@northsouth.edu)
• Mohammad Tushar Abdullah (tushar.abdullah.232@northsouth.edu)
• Asfee Bhuiyan Leen (asfee.leen.242@northsouth.edu)
• Sumaiya Tabassum Nimi (sumaiya.nimi@northsouth.edu)

Acknowledgements: Gemini 3.5 Flash-Lite generated the TRACE-BN teacher traces. GPT-4o-mini and DeepSeek-V4-Flash-0731 served as automated dual judges. AI tools were used for language editing.