The Gemma 4 Good Hackathon · Special Technology Track · Ollama prize
PathOS
Built by Dhairya Pandya, Shantanu Kharwar, and Rounak-agrwl
A histopathology research model with public Gemma weights and Ollama configuration. Training notebooks are available, but the claimed screening-to-report application pipeline and 88% accuracy were not independently confirmed.
I also participated in this hackathon with Memory Moment. This coverage was prepared independently after results were announced.
The idea worth stealing
SFT teaches what to say. GRPO teaches what not to say.
The team reports that PathOS v1 scored 26%, and 82% of its yes/no errors were not wrong answers but refusals: the base model’s instinct to hedge survived six hundred steps of supervised tuning. The team answered with four GRPO reward functions, one of which simply penalizes hedge language, plus an explicit answer tag so evaluation is exact. The team also reports that voting and retries improved a raw Ollama run by 55 points, then claims distillation reached 88% in one pass. Neither result was independently confirmed; the application pipeline is absent.
What we checked
Date checked: .
Public repository at the submission revision, Kaggle writeup, and linked artifacts; code was read, not executed. Deployment checks and limits are described below.
Labels apply to each finding and the evidence described, not to the project as a whole.
Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.
-
Four GRPO rewards, one against hedging
Verified
reward_no_hedge returns −1 when a regex finds cannot, unable, not sure, uncertain, or difficult to; the others score yes/no accuracy, the answer tag, and open answers. Vision layers train at 448 pixels. Source
-
v2’s 44 percent is in the notebook
Verified
The v2 notebook’s saved output reads “Yes/No accuracy: 22/50 (44.0%)”, on fifty questions. Source
-
88 percent and the five-tool pipeline
Not independently verified
The evaluation notebook has no outputs, generate_distillation_data.py is absent, and the analyze_patch-to-generate_report tool chain exists only in the writeup. Source
What we need: Executed evaluation outputs, the missing distillation script, and submitted application code with an end-to-end tool trace.
Under the hood
- The Modelfile bans doubt
- The system prompt reads NEVER hedge, lists cannot determine and uncertain as banned phrases, sets temperature 0.1, and stops generation at </answer>. A screening tool that cannot say “unclear” is a choice worth knowing.
- Weights were public before the deadline
- The Q8_0 GGUF and the PEFT adapter had their first Hugging Face commits on May 17; ollama pull dhairyapandya/pathos resolves to a 5.7 GB model.
Nuance
The repository is five notebooks, two Modelfiles, and a writeup; there is no application code, and the path from 26 to 88 percent rests on a 50-sample check we can see and a 100-sample one we cannot. The base model’s habit of hedging was trained out on purpose, which warrants scrutiny before interpreting screening outputs.
Sources
Last updated: . Editorial updates do not imply a new technical check.
Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.
Rounak-agrwl is the credit on the official submission and Kaggle profile; a full name was not confirmed.
Builder? Add context, request a correction, or ask for a re-check →