bangla-paraphrase-qwen2.5-qlora

TL;DR: A cautionary tale in this study — highest BLEU/ROUGE-L/BERTScore of all four models, but also the highest hallucination rate (22/100) under LLM-judge review. The clearest evidence in this project for why automated metrics alone are insufficient for evaluating paraphrase quality.

A QLoRA fine-tune of Qwen/Qwen2.5-1.5B-Instruct for Bengali paraphrase generation, trained on the BanglaParaphrase dataset. The LoRA adapter has been merged into the base weights — this is a complete, standalone model; no separate adapter loading is required.

This model is one of four architectures compared in a broader study, "Beyond BLEU: Evaluating Automated Metric Reliability for Bengali Paraphrase Generation" (GitHub repo), which benchmarks full fine-tuning vs. QLoRA and validates standard automated metrics (BLEU, ROUGE-L, BERTScore, PINC) against an LLM-as-a-judge. This model achieved the highest BLEU/ROUGE-L/BERTScore among all four, but also the most hallucination errors under LLM-judge review — see Limitations and Sample Outputs below.

Model Details

Trained on a single free-tier Colab T4 GPU with early stopping (43% of planned steps); despite the reduced training budget, this model achieved the highest surface-overlap metrics of the four — though not the most semantically reliable, see Evaluation Results.

Intended Use

Generating Bengali paraphrases of an input sentence — e.g., data augmentation for downstream Bengali NLP tasks, or research on low-resource paraphrase generation and parameter-efficient fine-tuning. Evaluated only on the BanglaParaphrase test distribution; not evaluated on other domains, dialects, or languages.

Out-of-Scope Use

• Not intended for languages other than Bengali, or for code-mixed/transliterated text.
• Not suitable for unreviewed automated use. This model had the highest hallucination rate of all four models studied (22/100 judged examples) — see Sample Outputs for a real failure case. Do not deploy outputs from this model without human review in any context where factual or semantic accuracy matters.
• Not a general-purpose Bengali instruction-following or chat model — although built on an instruction-tuned base, this checkpoint is fine-tuned specifically for sentence-level paraphrasing and has not been evaluated for other tasks.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "AvinabhDutta-Dev/bangla-paraphrase-qwen2.5-qlora"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")

text = "এই বাড়িতে সাধারণত কোনো ঝগড়াঝাঁটি হয় না।"
messages = [{"role": "user", "content": f"Paraphrase this Bengali sentence: {text}"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
output_ids = model.generate(inputs, max_new_tokens=160, do_sample=False)
print(tokenizer.decode(output_ids[0][inputs.shape[-1]:], skip_special_tokens=True))

Sample Outputs

Real outputs from this model on the held-out test set — one strong example and one genuine failure case, both from the study's LLM-judge evaluation, shown here deliberately for transparency rather than only showcasing favorable results:

Good paraphrase (judge: meaning 5/5, fluency 5/5):

Hallucination failure case (judge: meaning 1/5, fluency 2/5):

Judge's assessment: "The generated text distorts the original meaning completely and has poor grammar." The source describes officials taking cows from the speaker; the generated output garbles both the grammar and the direction of the action.

Evaluation Results

Evaluated on the full, held-out BanglaParaphrase test set (23,332 examples), using the same pipeline and test data across all four models in this study.

For context: a 70B-class zero-shot LLM baseline reported on this same dataset in prior work scores in the 10–12 BLEU range — this fine-tuned 1.5B model performs comparably despite being over an order of magnitude smaller, using only ~18.5M trainable parameters (~1.2% of the model's total).

LLM-as-a-Judge Validation

A 400-pair, four-model symmetric subset (100 examples per model, identical sources) was independently scored by Gemini 3.5 Flash-Lite on semantic meaning preservation and fluency (1–5 scale), with a binary meaning-reversal flag.

Error breakdown (out of 100 judged examples):

Related Models (this study)

• BanglaT5 (full fine-tune)
• IndicBARTSS (full fine-tune)
• TigerLLM-1B (QLoRA)

Limitations

• This model had the highest hallucination rate of all four models in the study (22/100 judged examples), despite scoring best on automated overlap metrics (BLEU, ROUGE-L, BERTScore). See Sample Outputs above for a real, representative failure case. This is a concrete illustration of why this study cross-validates automated metrics against LLM-judge scoring: strong surface-overlap metrics do not guarantee semantic reliability.

• Training-time validation loss was originally logged against a reduced 500-example subset for eval-frequency efficiency; a full-5K re-evaluation confirmed this was an accurate proxy (0.4113 vs. 0.4109).

• Trained on a 100K-example subset (not the full ~420K available training pairs) and evaluated with greedy decoding, for compute tractability.

• BLEU/ROUGE-L are known to correlate only weakly with actual paraphrase quality; this model is the clearest example of that gap in this study — high overlap scores, but the highest semantic-error rate under human-proxy review.

• Reproducibility: Trained on a fixed 100,000-pair subset of BanglaParaphrase with a documented train/val/test split (see processed_data/metadata.json in the GitHub repo); decoding was fully deterministic (greedy, do_sample=False), so generation outputs are reproducible given the same checkpoint and inputs.

Citation

If you use this model, please cite both the original Qwen2.5/BanglaParaphrase work and this study:

@misc{qwen2.5,
  title = {Qwen2.5: A Party of Foundation Models},
  url = {https://qwenlm.github.io/blog/qwen2.5/},
  author = {Qwen Team},
  month = {September},
  year = {2024}
}

@misc{banglaparaphrase2022,
  title={BanglaParaphrase: A High-Quality Bengali Paraphrase Dataset},
  author={Akil, Ajwad and others},
  year={2022}
}

See the project GitHub repository for the full paper, code, and citation for this study.

Framework Versions

• Transformers 5.13.1
• PEFT 0.19.1 (used during training; adapter merged into base weights for this release)

---

license: apache-2.0