Sankalp Thakur

Model · Reinforcement learning

industrial-edge-setpoint-kernels

Sankalp Thakur

Compact learned predictors and controllers for industrial setpoint control, tested against tuned PI and MPC on simulated plants: a two-pump pressure manifold, the Johansson four-tank system, and a thermal CSTR. It includes a C99 predictor for edge devices. All plants are synthetic, and the methods are adaptations of published papers, not reproductions. The offline FKDPP adaptation underperformed PI. Every outcome, including the losses, is kept in the reports. - The saved predictor is 12,484 bytes, with fixed-size inference. - The decision time is about 0.61 ms at the 95th percentile on an Apple ARM host. That excludes communication, PLC scans and admission checks. - The C99 predictor…

Model · Text generation

qwen3-06b-typed-decisions-cloud-pilot

Sankalp Thakur

This PEFT adapter was trained in private Kaggle kernel version 1 on 2026-09-30. It is a small, reproducible baseline for typed binary decisions. It is not Jev or a reproduction of Jev's undisclosed architecture. It does not establish a new small-model architecture or a general frontier-model breakthrough. The adapter contains only LoRA weights. Load it with the pinned base and the prompt/readout protocol in run.py; generic free-form generation is outside the measured setting. The downloaded adapter's SHA-256 is 827fd0d08c308a89df55343ad7045d9d5b9ac22394e4fa876bdef8c2a001c845. The base and adapted model were compared on the same records. The BoolQ confirmation set contains 256 test records…

Model

well-production-event-causal

Sankalp Thakur

Compact causal detectors for production events in real offshore well recordings from Petrobras 3W 2.0, tested on well groups they never saw during training. These are research models with their failures published. None is ready for deployment. The validation-selected primary is a 36-parameter logistic model. Two causal TCNs have 4,961 parameters each. Twenty-one of the covered episodes already had an alarm running when they began. In a later 15-method validation comparison, the selected logistic model caught 13 of 24 onsets within 300 s, 11 of them from one recording group. IBM's pretrained TSPulse caught none. - Every trained candidate and reference method, with the full report and…

Model · Image segmentation

controlled-diet-ogi-mask-reference

Sankalp Thakur

This is a small, reproducible test of whether compact CPU models can match plume masks in optical gas imaging (OGI) footage. The footage comes from baselab's Controlled Diet dataset: methane from in-vitro rumen fermentation in a laboratory, not industrial leaks. The result is negative. A simple spatial prior beat both learned models. Recording-level mean IoU against the publisher's generated masks, on 12 comparison frames from recordings the models never trained on: Thresholds and candidates were chosen on separate tuning recordings before the comparison ran. The models had unequal training budgets. All 11 candidates are kept in prepared/candidates/. - The scores measure agreement with weak…

Dataset · Image segmentation

cd-ogi-recording-mask-audit

Sankalp Thakur

Audit tables for a small gas-imaging study on baselab's Controlled Diet dataset, which is laboratory methane from rumen fermentation. Its 137 rows describe 42 source images from nine recordings. They are not images, chemical measurements or gas-free sessions. The spatial prior's comparison IoU is 0.138078, against 0.026416 for the tiny MLP and 0.022555 for logistic regression: both learned models lose. Empty cells are undefined metrics, not zeros. backgroundpixeloverpredictionratio measures pixel overprediction against weak masks. It is never a gas false-alarm rate. Models, weights and the full report are in by Mohamed G. Embaby, Toqi Tahamid Sarker, Amer AbuGhazaleh and Khaled R. Ahmed…

Dataset

johansson-delayed-sensor-control

Sankalp Thakur

Simulated trajectories of the Johansson quadruple-tank benchmark, for testing estimators and controllers when the lower-tank sensors arrive one sample late. There are 314,496 rows across four splits. These are simulations, not plant data. Keep these roles. Don't mix rows from shared trajectories at random. - At decision k, the allowed inputs are the public nominal regime, the input columns, and earlier rows of the same episode. - Condition names, seeds, and all result and label columns are never inputs. An applied result voltage becomes history only after it is applied. - labelnextmeasuredlower is y[k+1], which only arrives at decision k+2 because of the delay. - True tank levels, sensor…

Dataset · Tabular classification

3w-causal-real-group-audit

Sankalp Thakur

Decision-level and episode-level results from a well-event detection study on real Petrobras 3W 2.0 recordings, tested on well groups the detectors never saw in training. Use these tables to see how detectors transfer across wells. This is not a new sensor dataset; raw sensor values stay with the upstream source. - Scores, labels and provenance columns are for auditing. Never feed them to a predictor as inputs. - eventtimestampns is the dataset's nominal clock, not arrival or decision time. - Filename groups are anonymized public IDs, not authenticated well identities. The validation-selected detector covered 24 of 55 confirmation episodes and caught 3 of 12 observable onsets within 300…