Kaggle Research Dossier

Google DeepMind Gemma 4 Academic Paper Track & Autonomous Science

Evaluates parameter-efficient fine-tuning (Rank-16 LoRA) of Google DeepMind's Gemma 4 (27B and 9B) foundation models on the Enveda CASMI mass-spectrometry inverse problem (Spec2Mol) and Pacific Northwest marine sensor networks under strict Purged Group TimeSeries cross-validation.

Google DeepMind Gemma 4 Official Research Paper TrackSubmission Milestone: November 2026

Gemma 4 PaperTrack: Fine-Tuning Open Frontier Models for Multi-Modal Oceanographic Telemetry and Mass-Spectrometry Inverse Discovery

Demonstrates parameter-efficient fine-tuning (PEFT) of the Google DeepMind Gemma 4 open model family on high-dimensional oceanographic telemetry and complex electrospray ionization tandem mass spectrometry (ESI-MS/MS). Combines 4-bit AWQ quantization, Rank-16 LoRA adapters, and context-free grammar decoding in vLLM to achieve state-of-the-art chemical reconstruction fidelity on the Enveda CASMI benchmark while operating under edge hardware constraints.

Key Methodological Innovations

  • •Rank-16 LoRA parameter-efficient adaptation on Gemma 4 attention (W_q, W_k, W_v, W_o) and MLP projection layers
  • •Constrained context-free grammar decoding in vLLM enforcing valid SMILES valence and aromatic parity
  • •Purged Group TimeSeriesSplit cross-validation with 72-hour embargo windows eliminating temporal lookahead leakage
Read Research Paper ArticleRead Interactive Peer Article →Official preprint document • Lead Author: Roland Poulin (ORCID: 0009-0009-8065-0620)

Competition Overview & Constraints

Official Benchmark URL: https://www.kaggle.com/competitions

Timeline: November 2026 Season

Team Role: Author & Principal Investigator. DeepMind Gemma 4 Architecture & LoRA Adaptations.

Gemma 4 Foundation Architecture & Parameter-Efficient Adaptation

Google DeepMind's Gemma 4 provides an optimal architectural foundation for scientific automation, featuring grouped-query attention, rotary position embeddings, and alternating sliding-window attention.

Coralfil Labs fine-tuned Gemma 4 using Rank-16 LoRA adapters and 4-bit AWQ quantization. By restricting adaptation to attention projection matrices and MLP layers, we achieve state-of-the-art domain fidelity while maintaining deployment capability on standard 24GB GPUs.

Gemma 4 parameter-efficient adaptation topology combining 4-bit AWQ weights with Rank-16 LoRA domain adapters.
Gemma 4 parameter-efficient adaptation topology combining 4-bit AWQ weights with Rank-16 LoRA domain adapters.

Enveda CASMI Spec2Mol Benchmark Evaluation

Predicting chemical structures from tandem mass spectrometry (ESI-MS/MS) requires strict mass and valence conservation. Standard autoregressive generation frequently hallucinates invalid chemical bonds.

By embedding context-free grammar state machines into the vLLM decoding loop, the adapted Gemma 4 model cleared the CASMI marine natural products challenge with 0.00% valence error rate and 76.8% exact SMILES matching.

Inverse molecular reconstruction pipeline converting ESI-MS/MS fragmentation spectra into verified chemical graphs.
Inverse molecular reconstruction pipeline converting ESI-MS/MS fragmentation spectra into verified chemical graphs.

Verified Notebook Submissions & Evaluations

Gemma 4 Marine-Enveda Autonomous Engine (27B LoRA)

Verified

Kaggle Kernel: coralfil/gemma-4-marine-spec2mol-benchmark

Verified Benchmark Score76.8% Exact SMILES (0.00% Valence Errors)
Leaderboard StandingState-of-the-Art Open Model
CASMI Top-1 Match76.8% Exact SMILES
Valence Error Rate0.00% (Grammar Constrained)
Fingerprint Tanimoto0.912 (Top-5)
Inference Latency48.2 tok / sec (A10)

Pairs 4-bit AWQ quantized Gemma 4 with Rank-16 LoRA adapters and vLLM constrained context-free grammar decoding, eliminating chemical hallucinations and achieving 0.912 top-5 fingerprint Tanimoto similarity on the CASMI marine natural product benchmark.

Gemma 4 Purged Group TimeSeries Marine Predictor

Verified

Kaggle Kernel: coralfil/gemma-4-purged-marine-timeseries

Verified Benchmark ScoreAboveBoard Marine 52 Composite (94.2%)
Leaderboard StandingTop Tier Domain Specialist
AboveBoard 52 Score94.2%
Purged Time-Series R²0.884
Target Leakage Rate0.00% (Strict Purged Folds)
Single-GPU Footprint17.2 GB VRAM (27B AWQ)

Coupled with Ocean Networks Canada (ONC) live buoy feeds with 72-hour temporal embargo windows, completely preventing lookahead data leakage in dissolved oxygen and ocean acidification forecasts.