CBRS: Cognitive Blood Request System with Bilingual Dataset and Dual-Layer Filtering for Multi-Platform Social Streams

A. B. M. Alim Al Islam; Anik Saha; Anisa Binte Asad; K. G. Subarno Bithi; Mst. Fahmida Sultana Naznin; Zia Ul Hassan Abdullah

arxiv: 2604.16665 · v2 · submitted 2026-04-17 · 💻 cs.CL

CBRS: Cognitive Blood Request System with Bilingual Dataset and Dual-Layer Filtering for Multi-Platform Social Streams

Anik Saha , Mst. Fahmida Sultana Naznin , Zia Ul Hassan Abdullah , Anisa Binte Asad , K. G. Subarno Bithi , A. B. M. Alim Al Islam This is my paper

Pith reviewed 2026-05-10 08:07 UTC · model grok-4.3

classification 💻 cs.CL

keywords blood donation requestssocial media filteringbilingual datasetLoRA fine-tuninglanguage model parsinginformation extractionemergency responsedual-layer architecture

0 comments

The pith

A dual-layer system with a fine-tuned Llama model filters blood donation requests from social media at 99 percent accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents the Cognitive Blood Request System to automatically detect urgent blood donation messages on social media that typically get lost in high-volume communications. It creates a dataset of 11K messages in English, Bengali, and transliterated forms, including challenging negative examples to build robustness. A dual-layer filtering setup combined with a fine-tuned small language model achieves high accuracy in both filtering and parsing tasks while being efficient in resource use.

Core claim

CBRS achieves an impressive 99% accuracy and precision in filtering, surpassing benchmark methods. In the parsing task, our LoRA finetuned Llama-3.2-3B model achieves 92% zero-shot accuracy, surpassing the base model by 41.54% and exceeding the few-shot performance of GPT-4o-mini, Gemini-2.0-Flash, and other LLMs, while resulting in a 35X reduction in input token usage. This work lays a robust foundation for scalable, inclusive information extraction in time-sensitive, object-focused tasks.

What carries the argument

Dual-layer filtering architecture with a LoRA-finetuned Llama-3.2-3B model for bilingual blood request parsing on an 11K dataset with adversarial negatives.

If this is right

The system enables timely alerts for blood donation needs across multiple social platforms.
It reduces computational costs through lower token consumption compared to larger models.
The approach supports linguistic diversity including transliterated text common in social media.
High precision helps minimize false alerts in noisy online environments.
The open dataset supports further development of similar extraction systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This method could extend to extracting other time-critical signals such as disaster aid requests from social media.
Public release of the dataset and models allows community adaptation to new languages or additional platforms.
Linking the filter output directly to blood bank systems might speed up matching donors to specific needs.

Load-bearing premise

The curated 11K dataset with adversarial negatives is representative of real multi-platform social streams and that the dual-layer system plus fine-tuned model will maintain high accuracy and low false positives when deployed live without further retraining or platform-specific adjustments.

What would settle it

Running the deployed system on live social media streams for several weeks and measuring the actual false positive rate plus missed urgent requests through manual comparison.

Figures

Figures reproduced from arXiv: 2604.16665 by A. B. M. Alim Al Islam, Anik Saha, Anisa Binte Asad, K. G. Subarno Bithi, Mst. Fahmida Sultana Naznin, Zia Ul Hassan Abdullah.

**Figure 2.** Figure 2: Wordcloud of top keywords in CBRS dataset [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: Data sourcing process of CBRS: positive samples are collected from Facebook, EBDR-Twitter, and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: Dual-layered filtering and structured parsing architecture of CBRS, where raw messages undergo [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Two-layer DLF framework, with Layer 1 iden [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.** Figure 6: Comparison of parsing accuracy across differ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 7.** Figure 7: This figure illustrates the overall workflow [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8: Few-shot prompt for blood donation request parsing. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗

**Figure 9.** Figure 9: Few-shot prompt for generating adversarial negative samples for blood donation message classification. [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗

**Figure 14.** Figure 14: This figure shows demographic distribution [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗

**Figure 11.** Figure 11: This figure shows demographic distribution [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗

**Figure 16.** Figure 16: This figure shows demographic distribution [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗

**Figure 17.** Figure 17: This figure presents the results of a user study [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗

**Figure 18.** Figure 18: This figure shows Pearson Correlation Heatmap of User Feedback Metrics [PITH_FULL_IMAGE:figures/full_fig_p027_18.png] view at source ↗

read the original abstract

Urgent blood donation seeking posts and messages on social media often go unnoticed due to the overwhelming volume of daily communications. Traditional app-based systems, reliant on manual input, struggle to reach users in low-resource settings, delaying critical responses. To address this, we introduce the Cognitive Blood Request System (CBRS), a multi-platform framework that efficiently filters and parses blood donation requests from social media streams using a cost-efficient dual-layered architecture. To do so, we curate a novel dataset of 11K parsed blood donation request messages in Bengali, English, and transliterated Bengali, capturing the linguistic diversity of real social media communications. The inclusion of adversarial negatives further enhances the robustness of our model. CBRS achieves an impressive 99% accuracy and precision in filtering, surpassing benchmark methods. In the parsing task, our LoRA finetuned Llama-3.2-3B model achieves 92% zero-shot accuracy, surpassing the base model by 41.54% and exceeding the few-shot performance of GPT-4o-mini, Gemini-2.0-Flash, and other LLMs, while resulting in a 35X reduction in input token usage. This work lays a robust foundation for scalable, inclusive information extraction in time-sensitive, object-focused tasks. Our code, dataset, and trained models are publicly available at [https://github.com/aaniksahaa/CBRS](https://github.com/aaniksahaa/CBRS).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CBRS introduces a practical bilingual dataset and efficient parsing pipeline but its performance on real streams is not yet verified.

read the letter

The key takeaway is that this work creates a new 11K bilingual dataset for blood donation requests and demonstrates a dual-layer filtering system with a fine-tuned small LLM that cuts token usage significantly, but the strong performance numbers lack the evaluation details needed to confirm they hold up outside the lab. They curate messages in Bengali, English, and transliterated Bengali from social media, add adversarial negatives for better training, and build a pipeline that first filters relevant posts then parses them using LoRA on Llama-3.2-3B. The tuned model reaches 92% accuracy on parsing in zero-shot, a 41% lift over the base, and exceeds some bigger LLMs in few-shot while using 35 times fewer tokens. The filtering stage claims 99% accuracy and precision, better than benchmarks. Everything from code to models is released on GitHub. This is solid applied work for a specific real-world need in low-resource healthcare settings where manual apps don't reach everyone. The efficiency gain and public release stand out as practical steps forward. The soft spots center on generalization. All the headline results come from held-out parts of their curated dataset, and there's no information on test set construction, platform stratification, or any evaluation on live streams. The concern about whether the data matches real distributions is valid, and without that link the deployment claims are hard to assess. Minor gaps include limited error analysis and baseline details in the provided summary. This paper suits researchers building social media tools for urgent needs or working on low-resource NLP applications. Readers who want a starting point for similar extraction tasks or a template for dual-stage systems will get something usable from it. It deserves a serious referee because the dataset and pipeline are new and concrete, even if more testing is needed. I recommend sending it for peer review, with feedback focused on adding validation experiments and clearer method descriptions.

Referee Report

3 major / 1 minor

Summary. The paper introduces CBRS, a dual-layer filtering system for extracting blood donation requests from multi-platform social media streams (Bengali, English, transliterated Bengali). It curates an 11K dataset with adversarial negatives, applies a cost-efficient architecture, and fine-tunes Llama-3.2-3B via LoRA for parsing. The abstract claims 99% accuracy/precision on filtering (surpassing benchmarks) and 92% zero-shot parsing accuracy (41.54% lift over base model, exceeding few-shot GPT-4o-mini/Gemini-2.0-Flash while cutting tokens 35X), with public release of code, data, and models.

Significance. If the performance holds under real distributions, the work could enable faster, scalable donor matching in low-resource settings where social media is primary, advancing humanitarian NLP applications. The public dataset and models are a clear strength for reproducibility in bilingual social-stream extraction.

major comments (3)

[Abstract] Abstract: The 99% filtering accuracy/precision and 92% parsing accuracy are reported without any evaluation protocol, train/test splits, platform/language stratification, adversarial-negative sampling procedure, or error analysis. This is load-bearing for the central empirical claims, as the reader's report notes the absence of these details prevents assessing whether metrics reflect generalization or optimistic conditions on the curated set.
[Abstract] Abstract: The claim that the LoRA-tuned Llama-3.2-3B exceeds few-shot GPT-4o-mini, Gemini-2.0-Flash, and other LLMs by 41.54% (with 35X token reduction) lacks specification of the few-shot prompt format, example count, exact test-set construction, or whether the fine-tuned model was evaluated in a true zero-shot held-out regime. Without this, the comparative superiority cannot be verified.
[Abstract and Methods] Abstract and Methods: No cross-platform hold-out results, temporal splits, or live-stream deployment evaluation are described to support the claim that the dual-layer system plus fine-tuned model will maintain accuracy on real multi-platform streams without platform-specific retraining. The skeptic note correctly identifies this as the unsupported link between metrics and deployment utility.

minor comments (1)

[Abstract] Abstract: The dual-layer architecture is described only at high level; a brief enumeration of the two layers (e.g., rule-based then model-based) would improve immediate clarity even before full methods.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point by point below, providing clarifications from the manuscript where available and indicating revisions made to improve transparency.

read point-by-point responses

Referee: [Abstract] Abstract: The 99% filtering accuracy/precision and 92% parsing accuracy are reported without any evaluation protocol, train/test splits, platform/language stratification, adversarial-negative sampling procedure, or error analysis. This is load-bearing for the central empirical claims, as the reader's report notes the absence of these details prevents assessing whether metrics reflect generalization or optimistic conditions on the curated set.

Authors: We agree the abstract is too concise to stand alone on these points. The full manuscript details the evaluation protocol in Section 3 (including 80/20 stratified train/test splits by language and platform, 5-fold cross-validation for filtering, adversarial-negative sampling procedure in 3.3, and error analysis in Section 5). To make this accessible without requiring the full text, we have expanded the abstract to briefly reference the stratified splits, cross-validation, and adversarial sampling. revision: yes
Referee: [Abstract] Abstract: The claim that the LoRA-tuned Llama-3.2-3B exceeds few-shot GPT-4o-mini, Gemini-2.0-Flash, and other LLMs by 41.54% (with 35X token reduction) lacks specification of the few-shot prompt format, example count, exact test-set construction, or whether the fine-tuned model was evaluated in a true zero-shot held-out regime. Without this, the comparative superiority cannot be verified.

Authors: The manuscript specifies 5-shot prompting with the exact template in Appendix A, the held-out 20% test set (stratified, never seen during fine-tuning), and true zero-shot inference for the LoRA-tuned model on that set. Token counts were measured directly on the same inputs. We have added a clarifying clause to the abstract and a short methods paragraph to make the comparison protocol explicit. revision: yes
Referee: [Abstract and Methods] Abstract and Methods: No cross-platform hold-out results, temporal splits, or live-stream deployment evaluation are described to support the claim that the dual-layer system plus fine-tuned model will maintain accuracy on real multi-platform streams without platform-specific retraining. The skeptic note correctly identifies this as the unsupported link between metrics and deployment utility.

Authors: The current splits are stratified across platforms but do not include explicit cross-platform hold-out (train on one platform, test on another) or temporal splits; no live-stream deployment was performed. We acknowledge this weakens the direct claim of robustness without retraining and have added an explicit Limitations paragraph stating the stratified nature of the existing evaluation while noting the absence of cross-platform and deployment tests. Future work will address these. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical results on held-out data with no derivations or load-bearing self-citations

full rationale

The paper curates an 11K bilingual dataset, applies dual-layer filtering, and fine-tunes a LoRA-adapted Llama-3.2-3B model, then reports standard classification and parsing accuracies (99% filtering, 92% zero-shot parsing) on held-out test portions. No equations, parameter-fitting steps presented as predictions, or self-citation chains appear in the provided text. Claims rest on empirical evaluation against external baselines (GPT-4o-mini, Gemini, etc.) rather than reducing to inputs by construction. This is a typical applied ML paper whose central results are falsifiable on new data and therefore score at the low end of the scale.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on standard machine-learning assumptions about data representativeness and generalization rather than new theoretical constructs. No invented physical or mathematical entities are introduced.

free parameters (2)

Dual-layer decision thresholds
Thresholds or cutoffs used in the two filtering stages are not specified in the abstract but are typical free parameters in such pipelines.
LoRA hyperparameters
Rank, alpha, and learning-rate choices for fine-tuning the 3B model are free parameters whose values affect the reported 92% accuracy.

axioms (2)

domain assumption The distribution of real social-media blood requests matches the curated 11K dataset including adversarial negatives
Required for the 99% filtering accuracy to transfer to live multi-platform streams.
standard math Standard supervised fine-tuning assumptions (i.i.d. samples, stable optimization) hold for the LoRA-adapted Llama model
Invoked implicitly when claiming zero-shot generalization and token reduction.

pith-pipeline@v0.9.0 · 5609 in / 1679 out tokens · 35977 ms · 2026-05-10T08:07:09.686778+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

[1]

InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 14656–14672

Banglatlit: A benchmark dataset for back- transliteration of romanized bangla. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 14656–14672. Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Ari- vazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. InProceedings of the 60th Annual Meeting of the Association...

work page arXiv 2024
[2]

Do you request blood donations on social me- dia (e.g., Telegram, Discord, etc.)? (Almost always, Often, Sometimes, Seldom, Never)

work page
[3]

Did you usually receive timely responses to your blood donation requests before using BNet prior to October 23, 2024? (Almost always, Often, Sometimes, Seldom, Never)

work page 2024
[4]

How satisfied are you with the timely response of BNet in identifying potential donors be- tween October 23 and October 26, 2024, after integrating BNet into groups? (Very satisfied, Satisfied, Neither, Dissatisfied, Very dissatisfied)

work page 2024
[5]

After getting a response from BNet, have you successfully connected with a blood donor through BNet? (Almost always, Often, Sometimes, Seldom, Never)

work page
[6]

How easy do you find using BNet through slash command prompts? (Extremely easy, Very easy, Moderately easy, Slightly easy, Not at all)

work page
[7]

How intuitive is the user interface of BNet? (Extremely intuitive, Very intuitive, Moder- ately intuitive, Slightly intuitive, Not at all)

work page
[9]

At most how many blood donation seeking messages do you feel comfortable to receive from BNet per month? (1-5, 6-10, 11-15, 16-20, 21+)

work page
[10]

Do you find BNet more effective than exist- ing blood donation apps or methods you have used before? (Much better, Somewhat better, Stayed the same, Somewhat worse, Much worse, Not applicable- I have never used any app before)

work page
[11]

What challenges do you face in connecting with blood donors? How can these be over- come? (Open-ended response)

work page
[12]

What improvements would you suggest to make BNet better for requesters? (Open-ended response) For Donors:

work page
[13]

How many times have you donated blood in the past year? (Never, 1 time, 2 times, 3 times, 4 or more)

work page
[14]

Do you have trouble finding blood donation requests among a large volume of messages in social media groups? (Almost always, Often, Sometimes, Seldom, Never)

work page
[15]

How convenient is BNet in notifying you about blood donation requests in social media groups? (Extremely convenient, Very convenient, Mod- erately convenient, Slightly convenient, Not at all)

work page
[16]

How would you rate the overall functionality of BNet? (Excellent, Above Average, Average, Below Average, Very Poor)

work page
[17]

Do you find BNet more effective than existing blood donation apps or methods you’ve used before? (Much better, Somewhat better, Stayed the same, Somewhat worse, Much worse, Not applicable)

work page
[18]

What challenges do you face in connecting with blood requesters? How can these be over- come? (Open-ended response)

work page
[19]

Very satisfied,

What improvements would you suggest to make BNet better for donors? (Open-ended response) E Data Analysis To address existing gap of existing BDSs, we ask the following research questions in this work: • RQ1:How can a multi-platform bot be de- signed to seamlessly integrate with OSNs to accelerate donor response and broaden the donor network? • RQ2:How ca...

work page 2024

[1] [1]

InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 14656–14672

Banglatlit: A benchmark dataset for back- transliteration of romanized bangla. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 14656–14672. Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Ari- vazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. InProceedings of the 60th Annual Meeting of the Association...

work page arXiv 2024

[2] [2]

Do you request blood donations on social me- dia (e.g., Telegram, Discord, etc.)? (Almost always, Often, Sometimes, Seldom, Never)

work page

[3] [3]

Did you usually receive timely responses to your blood donation requests before using BNet prior to October 23, 2024? (Almost always, Often, Sometimes, Seldom, Never)

work page 2024

[4] [4]

How satisfied are you with the timely response of BNet in identifying potential donors be- tween October 23 and October 26, 2024, after integrating BNet into groups? (Very satisfied, Satisfied, Neither, Dissatisfied, Very dissatisfied)

work page 2024

[5] [5]

After getting a response from BNet, have you successfully connected with a blood donor through BNet? (Almost always, Often, Sometimes, Seldom, Never)

work page

[6] [6]

How easy do you find using BNet through slash command prompts? (Extremely easy, Very easy, Moderately easy, Slightly easy, Not at all)

work page

[7] [7]

How intuitive is the user interface of BNet? (Extremely intuitive, Very intuitive, Moder- ately intuitive, Slightly intuitive, Not at all)

work page

[8] [9]

At most how many blood donation seeking messages do you feel comfortable to receive from BNet per month? (1-5, 6-10, 11-15, 16-20, 21+)

work page

[9] [10]

Do you find BNet more effective than exist- ing blood donation apps or methods you have used before? (Much better, Somewhat better, Stayed the same, Somewhat worse, Much worse, Not applicable- I have never used any app before)

work page

[10] [11]

What challenges do you face in connecting with blood donors? How can these be over- come? (Open-ended response)

work page

[11] [12]

What improvements would you suggest to make BNet better for requesters? (Open-ended response) For Donors:

work page

[12] [13]

How many times have you donated blood in the past year? (Never, 1 time, 2 times, 3 times, 4 or more)

work page

[13] [14]

Do you have trouble finding blood donation requests among a large volume of messages in social media groups? (Almost always, Often, Sometimes, Seldom, Never)

work page

[14] [15]

How convenient is BNet in notifying you about blood donation requests in social media groups? (Extremely convenient, Very convenient, Mod- erately convenient, Slightly convenient, Not at all)

work page

[15] [16]

How would you rate the overall functionality of BNet? (Excellent, Above Average, Average, Below Average, Very Poor)

work page

[16] [17]

Do you find BNet more effective than existing blood donation apps or methods you’ve used before? (Much better, Somewhat better, Stayed the same, Somewhat worse, Much worse, Not applicable)

work page

[17] [18]

What challenges do you face in connecting with blood requesters? How can these be over- come? (Open-ended response)

work page

[18] [19]

Very satisfied,

What improvements would you suggest to make BNet better for donors? (Open-ended response) E Data Analysis To address existing gap of existing BDSs, we ask the following research questions in this work: • RQ1:How can a multi-platform bot be de- signed to seamlessly integrate with OSNs to accelerate donor response and broaden the donor network? • RQ2:How ca...

work page 2024