firelex/jeff

How firelex/jeff works

Jeff is a local, fine-tuned decision model: a caller submits a situation and questions, and the serving API returns calibrated option probabilities and choices. The broader repository supports assembling and authoring training data, training checkpoints, evaluating them, and exercising decisions in games. The serving and game call paths are sampled; other subsystem links reflect README-described capabilities and module names, since their implementation excerpts were not provided.

Main entry points: server.py Decision API · __main__.py Game runner · model.py Decision model · train.py Fine-tuning

Components 21

Decision serving

Decision API src/jeff/server.py
Model checkpoint
Calibrated answers

Model runtime

Decision model src/jeff/model.py
Input encoding src/jeff/encoder.py
Model decoding src/jeff/decoder.py

Training data and models

Public data prep src/jeff/data.py
Synthetic authoring src/jeff/sft_synthetic.py
Training schedule src/jeff/sft_pipeline.py
Fine-tuning src/jeff/train.py

Evaluation

Model evaluation src/jeff/evaluate.py
JevBench src/jeff/jevbench.py
LegalBench src/jeff/legalbench.py
CodeContests src/jeff/codecontests.py

Game experiments

Game runner src/jeff/games/__main__.py
Game client src/jeff/games/common.py
Game environments src/jeff/games/frogger.py
Pac-Man src/jeff/games/pacman.py
Doom src/jeff/games/doom.py

Other components

API caller
Open-model author

Connections 20

API caller → Decision API: submits questions
Decision API → Decision model: calls prediction
Decision API → Model checkpoint: loads weights
Decision model → Input encoding: encodes input
Decision model → Model decoding: runs inference
Decision API → Calibrated answers: returns probabilities
Public data prep → Fine-tuning: prepares examples
Open-model author → Synthetic authoring: supplies generations
Synthetic authoring → Training schedule: authors examples
Training schedule → Fine-tuning: schedules tranches
Fine-tuning → Model checkpoint: writes weights
Model evaluation → Model checkpoint: evaluates model
JevBench → Model checkpoint: benchmarks model
LegalBench → Model checkpoint: benchmarks model
CodeContests → Model checkpoint: benchmarks model
Game runner → Game environments: runs episodes
Game runner → Pac-Man: runs episodes
Game runner → Doom: runs episodes
Game runner → Game client: selects player
Game client → Decision API: requests decisions

More Python diagrams: kaarthik108/snowchat · kuandeng/lightgcn · raizamartin/gemini-code · fishcodetech/muteki · guokr/swagger-py-codegen

Drawn by GitDiagram on September 29, 2026 from the repository's file tree, README and source files. Read as Markdown ·