TCS & AMD AI Hackathon 2026

Gemmra Nidāna for Pharmacovigilance

AI-powered drug safety review with auditable reasoning. One fine-tuned model that automates 4 critical pharmacovigilance tasks — from 30 minutes per case to 10–20 seconds. Targeting the $8.3B → $15–18B PV market.

View on GitHub HuggingFace Model

0 Composite Score

0 Seriousness F1

0 Format Compliance

ICH E2A Compliant

AMD MI300X Optimized

Auditable Reasoning

gemmra — inference

Click terminal to take control

Four Tasks. One Model.

Gemmra performs all four critical ICSR assessment tasks with auditable reasoning — each one previously requiring 30 minutes of expert review.

Seriousness Classification

Validation Score 0.995 F1

Classifies adverse events against ICH E2A seriousness criteria — death, hospitalization, disability, life-threatening.

MedDRA Coding

Validation Score 0.667 Weighted

Maps adverse event narratives to standardized MedDRA Preferred Terms using BioDEX-derived medical terminology.

Labelling Assessment

Validation Score 0.801 F1

Determines whether the adverse event is already listed in the drug's approved labelling.

WHO-UMC Causality

Validation Score 0.986 Weighted

Assesses the causal relationship between drug and adverse event using WHO-UMC methodology.

See Gemmra think.

Real model outputs with extended thinking traces. Each assessment follows regulatory frameworks (ICH E2A for seriousness, WHO-UMC for causality) — producing the audit trail that pharmaceutical companies and agencies like FDA require.

T1 Seriousness ACTEMRA 445 tokens

SERIOUS: YES

Patient: 69-year-old female · Events: Cardiac arrest, Pulmonary embolism, Acute kidney injury, + 11 more AEs

Thinking Trace

Pharmacovigilance expert.
Assess seriousness per ICH E2A criteria.
Death (DE), Life-threatening (LT), Hospitalization (HO), Disability (DS), Congenital anomaly (CA).

Patient: 69-year-old female.
Drug: ACTEMRA.
Adverse events: Acute kidney injury, ALT increased, AST increased, Blood glucose decreased, Blood potassium increased, CRP increased, Cardiac arrest, Fibrin D dimer increased, Haemodialysis, Inflammatory marker increased, Platelet count decreased, Pulmonary embolism, Serum ferritin increased, Therapy non-responder.

Outcome: "The patient did not survive the clinical episode." -> Death (DE).
Outcome: "The patient suffered permanent functional limitation." -> Disability (DS).
Clinical severity: Cardiac arrest and pulmonary embolism are life-threatening conditions. -> Life-threatening (LT).
Treatment: Haemodialysis suggests hospitalization or intensive care. -> Hospitalization (HO).

DE (Death): Yes. The patient died.
LT (Life-threatening): Yes. The clinical situation was critical.
HO (Hospitalization): Yes. The severity of the event required hospitalization.
DS (Disability): Yes. Permanent functional limitation is noted.
CA (Congenital anomaly): No. Not applicable.

The case meets multiple seriousness criteria (DE, LT, HO, DS).
Final classification: Serious.

Structured Output

SERIOUS: YES
Criteria met: DE (Death), LT (Life-threatening), HO (Hospitalization), DS (Disability)
Rationale: The clinical outcome meets multiple seriousness categories, confirming serious classification.

Why this matters: Checks all 5 ICH E2A criteria against 14 adverse events. Identifies 4 criteria met. Zero false positives across 1,027 T1 evaluation samples.

T4 Causality YESCARTA 541 tokens

WHO-UMC: Possible

Patient: 31-year-old female · Events: Cytokine release syndrome (5 days after treatment)

Thinking Trace

Pharmacovigilance expert.
Extract evidence and assess causality using WHO-UMC criteria.
31-year-old female. YESCARTA. Follicular lymphoma.
Cytokine release syndrome. 5 days.

Temporal Relationship: The 5-day interval between drug start and event onset is within the expected window for this type of reaction.
Dechallenge: No dechallenge information is available — this key criterion cannot be evaluated.
Rechallenge: The drug was not reintroduced, so rechallenge evidence is absent.
Confounders: No obvious confounders identified between the drug indication and the adverse event.
Alternatives: The adverse event is distinct from the treated condition, reducing confounding risk.
Concomitants: No concomitant medications were reported — the suspect drug is the sole agent.

Temporal: Strong (5 days).
Dechallenge: Unknown.
Rechallenge: Not performed.
Confounders: None.
The evidence is suggestive but not conclusive. The temporal link is strong, but the absence of dechallenge/rechallenge data limits the assessment to 'Possible'.

Structured Output

WHO-UMC Causality: Possible
Evidence:
  - Temporal: Strong (5-day interval, within expected window)
  - Dechallenge: Unknown (key criterion cannot be evaluated)
  - Rechallenge: Not performed
  - Confounders: None identified
  - Alternatives: AE distinct from treated condition
  - Concomitants: Sole agent

Why this matters: 6-dimension WHO-UMC evidence extraction. Correctly limits to "Possible" due to missing dechallenge data. 95.5% exact match across 794 T4 evaluation samples.

Base model vs. Gemmra.

Same model, same hardware. The difference is data-first fine-tuning — plus auditable reasoning traces and 100% format compliance.
Base model evaluated with format-friendly prompts & thinking disabled (250 samples). Gemmra evaluated on 3,645 decontaminated samples with zero prompt engineering.

Task	Metric	Base Gemma 4†	Gemmra (SFT)	Improvement	Samples
T1 Seriousness	F1 Score	0.977	0.995	+1.8%	1,013
T2 MedDRA Coding	Weighted	0.311	0.667	+114%	845
T3 Labelling	F1 Score	0.782	0.801	+2.4%	995
T4 Causality	Weighted	0.845	0.986	+16.7%	792
Composite Score	Average	0.729	0.862	+18.2%	3,645
Format Compliance	All tasks	N/A‡	100%	—	3,645

💡

Key finding: The base model achieves 0.729 composite only with hand-crafted format prompts and thinking disabled. Gemmra achieves 0.862 with zero prompt engineering, adds auditable thinking traces, and hits 100% format compliance — the gap between "works in a lab" and "works in production."
†Base evaluated with task-specific format instructions & enable_thinking=False ‡Format compliance not measured for base model

The data-first debugging journey.

We never changed the model architecture. Every improvement came from fixing the training data.

v1 Initial training

Composite 0.620

First SFT checkpoint (0.23 epoch). T1/T4 already strong — model gets seriousness and causality right early.

T1: 1.000 T2: 0.134 T3: 0.387 T4: 0.960

v5 +2.1× T2

Composite 0.705

Found abstract[:500] truncation hiding ground truth from 92% of T2 training data. One line fix → T2 jumped 2.1×.

T1: 1.000 T2: 0.558 T3: 0.286 T4: 0.975

v6 +3× T3

Composite 0.859

Forensic audit found 62% NO / 38% YES label imbalance in T3. Rebalancing to 50:50 restored T3 from 0.286 to 0.884.

T1: 1.000 T2: 0.572 T3: 0.884 T4: 0.980

Final Shipped ✓

Composite 0.862

Validated across 3,645 decontaminated samples. T2 improved further at scale (0.572→0.667). Composite confirmed at 0.862.

T1: 0.995 T2: 0.667 T3: 0.801 T4: 0.986

Three approaches. One conclusion.

We explored SFT, weight interpolation, and reinforcement learning. Each proved the same thing: data quality is the bottleneck, not the algorithm.

SFT

Shipped

0.862 composite

Format 100%

Reasoning 45 words

Collapsed

Supervised fine-tuning with completion-only loss. Data quality drove all gains — 4 iterative data fixes over 7 days.

Primary Model

WiSE-FT

Alternative

0.825 composite

Format 99%

Reasoning 400+ words

Deep reasoning

Weight interpolation explored to recover reasoning depth. α=0.9 trades ~4% accuracy for 10× richer thinking traces — a research contribution, not the primary model.

DAPO/GRPO

Validated Ceiling

0.828 composite

Format 100%

Reasoning 83–279 tokens

Plain text

DAPO (a frontier GRPO variant) on SFT gained only +0.003 composite, then reward variance collapsed. Both approaches prove SFT captured all learnable signal.

WiSE-FT α Sweep — Exploring the Accuracy vs. Reasoning Trade-Off

Blending fine-tuned weights with pre-trained weights at different α values. Higher α = more SFT format, lower α = more base model reasoning.

α	Composite	T1 F1	T2	T3 F1	T4	Format	Reasoning
1.0 (SFT)	0.862	0.995	0.667	0.801	0.986	100%	45 words
0.9	0.825	0.985	0.532	0.852	0.930	99%	400+ words
0.8	0.782	0.971	0.416	0.825	0.915	99.5%	400+ words
0.7	0.703	0.939	0.286	0.867	0.720	99.5%	400+ words

α=1.0 (pure SFT) is the primary model with highest accuracy. α=0.9 recovers deeper clinical reasoning at a ~4% accuracy cost — useful when audit trail depth matters more than raw metrics. T3 improved from 0.801→0.852 as base model clinical knowledge mixed in.

End-to-End Pipeline.

Three interconnected pipelines — from raw FDA data to validated clinical AI.
Every step is reproducible, open source, and documented.

FDA FAERS + BioDEX + OnSIDES → 32,355 training pairs across 4 task types

Gemmra Data Pipeline — FAERS, BioDEX, OnSIDES ingestion through cleaning, prompt building, and decontamination

bf16 LoRA SFT on Gemma 4 31B-it → 95 GB VRAM on AMD MI300X (192 GB HBM3)

Gemmra Training Pipeline — bf16 LoRA SFT on Gemma 4 31B using AMD MI300X 192GB HBM3

3,645 decontaminated samples → composite 0.862 across T1–T4 pharmacovigilance tasks

Gemmra Evaluation Pipeline — 3,645 decontaminated samples across T1-T4 pharmacovigilance tasks

Built for Production.

How Gemmra aligns with enterprise pharmaceutical operations and AMD hardware strategy.

TCS Life Sciences

TCS Life Sciences is a Leader in pharmacovigilance operations (Everest Group PEAK Matrix), serving top global pharmaceutical companies. Drug safety case processing is a core operational challenge for these clients — averaging 100K+ cases/year per company. Gemmra directly addresses this by automating the four most time-consuming pharmacovigilance assessments.

~2,000× Cost reduction per case

~100–180× Faster than manual review

100% Format compliance

AMD MI300X Advantage

Our training requires 95 GB VRAM for bf16 LoRA across all linear layers of a 31B model. This is physically impossible on any single NVIDIA GPU (A100/H100 max at 80 GB). AMD MI300X's 192 GB HBM3 is the enabling technology — zero quantization means higher quality gradients and a better final model.

95 / 192 GB Training VRAM (bf16)

61 / 192 GB Inference VRAM (31.8%)

1.9 hrs Full SFT training time

Data Challenges We Solved

We claim "data-first debugging" — here's the evidence:

MedDRA is proprietary → Engineered PT training from BioDEX open biomedical literature + FAERS reaction terms

FDA redacts doctor narratives → Built structured prompts from remaining FAERS fields (demographics, drug info, outcomes)

BioDEX truncated at 500 chars → Fixed one line → 2.1× improvement on MedDRA coding (92% of training data was corrupted)

No train/eval decontamination → MeditronFO-inspired content-hash splitting across 3,645 eval samples

No drug-label ground truth → Integrated OnSIDES (drug label side effects database) for T3 labelling assessment ground truth

Engineered on AMD MI300X.

Full bf16 precision, zero quantization, LoRA r=64. Every architectural choice leverages AMD's 192 GB HBM3 advantage.

bf16 LoRA Training

Full bf16 precision with LoRA rank 64 on MI300X. Zero quantization = zero gradient noise. LoRA r=64

Extended Thinking Mode

Gemma 4's native thinking mode generates visible reasoning traces — the audit trail regulators require. enable_thinking=True

Data-First Debugging

Our biggest gain came from fixing one data pipeline bug, not hyperparameter tuning. Data quality > model tricks. 2.1× improvement

Exact Match

Synonym Match

Fuzzy Match

SOC Match

Gemmra-Bench v1.0

Novel 4-level hierarchical scoring: exact → synonym → fuzzy → SOC match. 3,645 decontaminated evaluation samples across 4 PV tasks. Open Benchmark

MI300X Inference

8.5s per case · 6.7 tok/s · 500ms TTFT · 61 GB / 192 GB VRAM (31.8%)

Published Negative Results

RL (DAPO/GRPO) on SFT gained only +0.003 composite, then reward variance collapsed. WiSE-FT trades 4% accuracy for richer reasoning. Both prove SFT captured all learnable signal. Scientific Rigor

Team 2930 · FINETUNING_005