WHY THIS
EXISTS.
Fine-tuning a local model still feels like cluster work, while many experiments get promoted because one demo looked good. The missing piece is a repeatable path from curated examples to a measured, reversible adapter.
OSS / ACTIVE · AI & models
// Make the score earn trust
LLM Gym
A local training and evaluation workbench for small language models, LoRA adapters, human feedback and measurable promotion gates.
LAST PUSH…syncing with GitHub
LANGUAGE…primary repository language
RELEASENO FORMAL RELEASElatest public GitHub release
LICENSENOT DECLAREDrepository metadata
WHY THIS
THE USEFUL
LOOP.
LLM Gym keeps training pools, LoRA definitions, hardware-aware backends, verification and promotion in one local dashboard. Adapters move forward only after explicit checks and human gates.
03 // WHAT_SHIPPEDWHAT'S
WHAT'S
INSIDE.
ADAPTER WORKBENCH
Define one narrow skill, curate its JSONL pool and train a small swappable LoRA instead of another monolith.
HARDWARE-AWARE QUEUE
MLX, PEFT and simulate backends share a single queue that avoids overlapping jobs and memory blowups.
VERIFY BEFORE PROMOTE
Acceptance prompts, local judges, pool forecasts and DPO feedback turn promotion into a measurable decision.