Moonshot released Kimi K3 with inference code only. I rebuilt the training pipeline from scratch, fixed the four places where the released code cannot train, and trained a 1.27B test model with it.
Research intern at IIT Kharagpur Kolkata, India
Arnesh Banerjee
Machine learning & software engineering
I’m a final-year computer science student. I work on medical imaging, multi-omics cancer modelling and LLM alignment, with research at IIT Kharagpur, Jadavpur University and NJIT. On the side, I rebuilt the training pipeline for Kimi K3 and contribute to open source.
Experience03
IIT Kharagpur
May 2026 – now
Summer research intern, with Dr. Sourangshu Bhattacharya
A multi-omic network for head-and-neck cancer, where DNA methylation
alone tells oral from non-oral tumour origin at
0.937 accuracy.
- Benchmarked an improved MoXGATE multi-omic network (0.827 to 0.927 accuracy) against LightGBM on 6 TCGA data types from 528 head-and-neck cancer patients, across all 63 data-type combinations (2,040 models per pipeline).
- Showed that DNA methylation alone predicts oral vs. non-oral tumour origin at 0.937 accuracy.
- Explained the model with TreeSHAP: 87% of its decision traces to methylation probes in the developmental HOX and PITX2 genes, and a 50-probe signature keeps 0.935 accuracy.
- Ruled out HPV and hospital confounds, with 16 of 16 HPV-negative laryngeal tumours correct.
Jadavpur University
Nov 2025 – May 2026
Research intern, CMATER Lab, with Prof. Debotosh Bhattacharjee
A hybrid CNN-Transformer that segments breast thermograms at
0.9716 Dice.
- ResNet-34 encoder, ASPP, a Transformer bottleneck and SE-gated skips, trained with a differentiable Chan-Vese loss: 0.9716 Dice and 0.9463 IoU on DMR-IR (357 thermograms, 119 patients), patient-stratified 5-fold.
- Benchmarked against Attention U-Net, UNet++, DeepLabV3+ and TransUNet on five metrics. All reach about 0.97 Dice (p > 0.05), which shows the annotations are the ceiling.
- Added robustness tests (label noise, augmentation, smaller training sets) and explainability (Grad-CAM, attention maps, MC-dropout uncertainty) for clinicians.
New Jersey Institute of Technology
Jun – Nov 2025
Summer research intern, Learning-based Decision Making Lab, with Dr. Arnob Ghosh
A semantic safety cost model for LLM alignment that agrees with human
safety judgments 92% of the time.
- Built a 2,500-pair safety dataset (benign, context-sensitive and jailbreak prompts) and fine-tuned the last six layers of LLaMA-2-7B-chat into the cost model.
- Scored 0.91 to 0.96 on XS-Test, against 0.07 to 0.32 for Safe-RLHF, so it does not refuse on keywords alone.
- The trained policy flags unsafe responses 5× more efficiently than Safe-RLHF and wins 60% of head-to-head comparisons (+70 Elo) against SACPO. Co-author of the preprint.
Projects03
A Zed extension for Jupyter notebooks. A language server checks
.ipynb files, and code actions run cells on real kernels, with
plots and tables inline.
LLMs debate a topic over several rounds as anonymous agents. Each agent gets a fresh random code every round, so the judge never learns which model wrote what.
Research04
Recursive and Wrapper-Based Feature Selection for Breast Cancer Diagnosis and Prognosisoral
2026An Intelligent Weakly Supervised Framework for Breast Thermography Segmentation Using Hybrid CNN–Transformer Networks
in prepOpen source
Merged pull requests to projects such as CPython, Rust, Zed, OpenCV and Kubescape.
A year on GitHub
Latest merged
Writing02
Amazon ML Challenge 2026 in a self-imposed 24 hours
Sep 2026Matching 10 million noisy business records to 1.7 million businesses across the US, India and an unseen France, for 0.988216 F0.5.
Kimi K3 for All
Aug 2026Rebuilding the training pipeline Moonshot never released, and the small model it produced along the way.
Guestbook
Loading the guestbook…