skblv/gemma-3-27b-it-lora-json-cholect50-verbs

Gemma 3 27B-it LoRA (JSON) — CholecT50 verb recognition

Generative LoRA adapter on google/gemma-3-27b-it, fine-tuned to emit a JSON action list for CholecT50 frames.

This is the “Gemma 3 27B + LoRA (JSON)” row on the Recommendations tab of the SDSC × Chicago Booth surgical video understanding leaderboard.

Prompt example

[surgical frame]

Identify all surgical actions currently being performed in this laparoscopic cholecystectomy video frame.

Valid action names:
- grasp
- retract
- dissect
- coagulate
- clip
- cut
- aspirate
- irrigate
- pack
- idle

Return your answer as a JSON object with this exact format:
{"current_actions": ["action_1", "action_2"]}

Use ONLY the exact action names from the list above.
If no action is being performed, return: {"current_actions": ["idle"]}
Return ONLY valid JSON, nothing else.

Files

• adapter_model.safetensors / adapter_config.json — LoRA (r=16, alpha=32) on q/k/v/o_proj and gate/up/down_proj
• training_config.json — training hyperparameters

This is a generative JSON adapter, not a classification-head checkpoint.

Training

• 2 epochs, lr 1e-4, effective batch size 4, seed 42
• 76,893 fit frames / 4,047 monitor frames / 19,923 validation frames

Evaluation

Full CholecT50 verbs validation split (95% bootstrap CI; 5 parse failures on 19,923 frames):

Metric	Value
Exact match	60.0% (59.3–60.7)
Micro-averaged F1	76.6% (76.1–77.0)

References

• Skobelev, K., Fithian, E., Baranovski, Y., et al. A Comparative Study in Surgical AI: Potential and Limitations of Data, Compute, and Scaling. arXiv:2603.27341, 2026.
• Dataset: Nwoye, C. I., Yu, T., Gonzalez, C., et al. Rendezvous: Attention Mechanisms for the Recognition of Surgical Action Triplets in Endoscopic Videos. Medical Image Analysis, 78, 102433 (2022).

Limitations

Research baseline only. Not a medical device.