Question Solved

What does a typical data management plan look like for a SERB project?

Back to Forum
Growing 195 pts 0 followers
Manipal Institute of Technology · Posted
Earth & Climate Science Urban Planning

Background

I am a second-year PhD student working on multilingual NLP at NITK. My research involves building named entity recognition systems for code-mixed Telugu-English text, but I am struggling with the fine-tuning process for large language models.

The problem

I have around 8,000 annotated sentences, which I know is relatively small for fine-tuning something like XLM-R. When I fine-tune with default settings, I get reasonable results on validation but the model overfits badly by epoch 5.

What I have tried

  • Reduced learning rate to 1e-5
  • Added dropout at 0.3
  • Early stopping with patience=3
  • Tried LoRA fine-tuning as an alternative

Has anyone dealt with similar low-resource scenarios? Is LoRA actually recommended for NER tasks, or is full fine-tuning with heavy regularization the better path?

Any references to papers or GitHub repos with similar setups would be very helpful.

Sign in to interact

Sign in to join the discussion.

28 Replies

Oldest Top
0
Fatima Sheikh Growing 95 pts · Accepted answer

The ethics review situation in Indian universities is actually more complex than most people realize.

There is no uniform national framework equivalent to the US Common Rule. Each institution has its own Institutional Ethics Committee (IEC) or Institutional Review Board (IRB).

For anonymous survey research with no identifiable data and no vulnerable populations: most Indian IECs would classify this as exempt, but 'exempt' still needs a formal determination from the IEC.it is not self-certifying.

My practical advice: go to your institution's research ethics office now, describe what you did, and request an exempt determination in writing. Better to have documentation than to discover the issue during thesis review.

0
Replying to Fatima Sheikh
Revathi Chandrasekaran Active 410 pts · Accepted answer

This question comes up a lot. The answer really depends on your specific field and what your committee values.

0
Replying to Fatima Sheikh
Fatima Sheikh Growing 95 pts · Accepted answer

I am in a very similar situation. Would you be willing to share the outline of your PMRF proposal? Not the content.just the section headings and approximate word allocation.

0
Disha Malhotra Starting 8 pts · Accepted answer

I have successfully received PMRF (lateral entry, second year). Happy to share what worked in my proposal.

On the research proposal:

  • 60–70% technical approach, 30% context and impact
  • Preliminary results absolutely help. Coursework projects are fine if genuinely relevant
  • The societal impact section should be substantive.PMRF is a national initiative; reviewers want to see why this matters

On citations: Yes, cite your MTech thesis. Framing your PhD as an extension of prior work you did shows continuity of thought.

0
Meghana Rao Growing 170 pts · Accepted answer

Great question.I went through something very similar in my second year.

The key insight for me was that LoRA is actually quite well-suited for NER tasks, especially in low-resource settings. I would recommend:

  1. Use LoRA with r=8 or r=16.don't go higher for 8K samples
  2. Apply LoRA to attention layers only, not the feed-forward layers
  3. Use a cosine learning rate schedule with warm-up (10% of steps)

For Telugu-English code-mixed NER specifically, you might also look at MuRIL.it is pretrained on Indian language data and often outperforms XLM-R on Indic tasks even with less fine-tuning data.

0
Replying to Meghana Rao
Quamar Ahmed Growing 65 pts · Accepted answer

If you are not already using a reading group format with your students, I would strongly recommend it. Forces everyone to articulate what they read.

0
Replying to Meghana Rao
Kiran Yadav Growing 130 pts · Accepted answer

Strongly agree with the point about choosing venues carefully. One paper in NeurIPS > five papers in obscure workshops.

0
Chetan Jain Growing 100 pts · Accepted answer

Great question.I went through something very similar in my second year.

The key insight for me was that LoRA is actually quite well-suited for NER tasks, especially in low-resource settings. I would recommend:

  1. Use LoRA with r=8 or r=16.don't go higher for 8K samples
  2. Apply LoRA to attention layers only, not the feed-forward layers
  3. Use a cosine learning rate schedule with warm-up (10% of steps)

For Telugu-English code-mixed NER specifically, you might also look at MuRIL.it is pretrained on Indian language data and often outperforms XLM-R on Indic tasks even with less fine-tuning data.

0
Admin NITK · Accepted answer

Overleaf + GitHub integration is underrated. You get version control and collaboration in one place.

0
Replying to Admin NITK
Royal Dsouza Starting 30 pts · Accepted answer

For longitudinal data with missing values, mixed-effects models handle this more gracefully than imputation in most cases.

0
Replying to Admin NITK
Nandita Ghosh Distinguished 720 pts · Accepted answer

I found that writing a paper abstract first (before the paper) and then reverse-engineering the paper from the abstract helped with clarity enormously.

0
Meena Sharma · Accepted answer

I can answer your question about IEEE Transactions review timelines from my experience submitting to IEEE TNNLS.

For Q1 IEEE journals, expect:

  • Initial desk review: 2–4 weeks
  • Full peer review: 3–6 months
  • Major revision decision: another 2–3 months for re-review

As for code and data, most IEEE journals now request code upon acceptance, not submission. Regarding your 2.3% improvement.that is borderline. You will need very strong experimental rigor: multiple runs, confidence intervals, and significance tests.

0
Priya Nair · Accepted answer

I think you are conflating two separate problems here. The publish-or-perish culture creates pressure, yes.but the solution is not to criticize researchers who are navigating an unfair system. The real problem is with evaluation committees who treat publication count as a proxy for research quality. Fix the evaluation, and the incentives change.

Also worth noting: open review has been implemented in some fields and has its own problems. It can disadvantage early-career researchers who fear retaliation from senior colleagues they have reviewed critically.

0
Replying to Priya Nair
Pooja Nambiar Growing 115 pts · Accepted answer

When you say 'multiple runs with confidence intervals', do you mean running the full training pipeline multiple times with different random seeds and reporting mean ± std? Or something more statistically rigorous?

0
Replying to Priya Nair
Quamar Ahmed Growing 65 pts · Accepted answer

Could you clarify what you mean by results-blind review? I have heard of it but never understood how it works in practice. Wouldn't reviewers need to see results to evaluate whether the methodology is sound?

0
Replying to Priya Nair
Wajid Khan Growing 110 pts · Accepted answer

Marking this as the accepted answer. Incredibly helpful and specific. Really appreciate you taking the time.

0
Sunil Bhattacharya Active 390 pts · Accepted answer

The Fulbright SoP is very different from a typical PhD SoP. It needs to focus on cultural exchange and diplomacy, not just research.

0
Replying to Sunil Bhattacharya
Royal Dsouza Starting 30 pts · Accepted answer

This is exactly what I needed. Thank you so much.I have been going back and forth on this for weeks and this clears it up completely.

0
Replying to Sunil Bhattacharya
Neha Agarwal Growing 90 pts · Accepted answer

This is exactly what I needed. Thank you so much.I have been going back and forth on this for weeks and this clears it up completely.

0
Gopal Naidu · Accepted answer

Thank you for this detailed answer! A quick follow-up: when you mention cosine learning rate with warm-up, are you warming up the LoRA parameters specifically or the entire model including the frozen backbone?

0
Nandita Ghosh Distinguished 720 pts · Accepted answer

If you are not already using a reading group format with your students, I would strongly recommend it. Forces everyone to articulate what they read.

0
Anita Rao · Accepted answer

Seconding the recommendation for Zotero. Game changer for managing references across multiple projects.

0
Replying to Anita Rao
Arjun Venkatesan Growing 195 pts · Accepted answer

Thank you! The reference to MuRIL is particularly useful.I had not considered it as an alternative to XLM-R for Indic languages.

0
Replying to Anita Rao
Sunil Bhattacharya Active 390 pts · Accepted answer

Really appreciate you taking the time to write this out in detail. This is going straight into my research notes.

0
Chetan Jain Growing 100 pts · Accepted answer

This is really helpful context. One more question: for the ethics committee approval, does the institution where the data is collected or the institution where the analysis is performed need to be the primary ethics approver?

0
Replying to Chetan Jain
Hemant Patwa · Accepted answer

Really appreciate you taking the time to write this out in detail. This is going straight into my research notes.

0
Replying to Chetan Jain
Lakshmi Devi Growing 75 pts · Accepted answer

This is exactly what I needed. Thank you so much.I have been going back and forth on this for weeks and this clears it up completely.

0
Replying to Chetan Jain
Aditya Sharma Starting 45 pts · Accepted answer

Much appreciated. I will restructure my proposal along these lines. The point about societal impact being substantive is something I would have gotten wrong.