How christy62/llm-g-2 works
This repository is a research prototype for reconstructing positron tracks by translating detector-hit sequences into decay-vertex and momentum parameters. Its visible workflow spans simulated decay generation, helix propagation to detector hits, optional strip digitization, conversion of event data into training examples, and LLM fine-tuning. Separate conversion scripts cover alternate hit formats. The README describes model-generated forward predictions and track-match results, but places inference in an external Colab notebook. Some data handoffs and the inference implementation are not established by the sampled source, so those connections are omitted.
Main entry points: generate_hits.py Hit propagation · fine_tune_llm.py LLM fine-tuning · generate_decay_details.py Decay generation · plot_helix.py Track plotting
Components 14
Event simulation
- Decay generation
generate_decay_details.py - Hit propagation
generate_hits.py - Scattering propagation
hit_details_mcs.py
Hit encoding
- Strip mapping
strip_id.py - Strip tokenization
vane_sensor_strip.py - MCS dataset conversion
csv_to_json.py - Random-hit conversion
new_json.py - Geant4 conversion
g4_json.py
LLM training
- LLM fine-tuning
fine_tune_llm.py - Trained model
fine_tune_llm.py
Validation and results
- Forward event format
gen_event.py - Track plotting
plot_helix.py
Other components
- Researcher
- Inference notebook
Connections 6
- Researcher → Decay generation: generates events
README.md - Researcher → Hit propagation: generates hits
README.md - Researcher → LLM fine-tuning: trains model
README.md - LLM fine-tuning → Trained model: produces
fine_tune_llm.py - Trained model → Inference notebook: inference available
README.md - Hit propagation → Track plotting: visualizes tracks