A toolkit for exploring, measuring, comparing, and improving language models through interactive and reproducible experiments.
- Browser demo: https://simulatte.world
- Architecture map: docs/ARCHITECTURE.md
- Full docs index: docs/README.md
GAMMA is deliberately a family name rather than one rigid expansion. Each backronym names a real project surface or goal:
| Context | Reading | What it names |
|---|---|---|
| Interactive game | Game Analyzing Model Methods Attentively | Predicting model continuations and inspecting the choices behind them. |
| Comparison and benchmarking | Guessing Alternative Model Mechanics Analytically | Comparing models, engines, decoding policies, and performance under recorded contracts. |
| Learning and interpretability | Grasping Attention Mechanism Mysteries Accessibly | Making logits, tokenization, attention-capable backends, and generation mechanics understandable. |
| SAME-R research | Generalized Approach Matching, Measurement, and Attribution | Testing different methods under matched evaluation and determining what caused an observed improvement. |
These are contextual readings, not separate product names. GAMMA remains the
stable project and command identity.
| Surface | Purpose | Starting point |
|---|---|---|
| Game and chat | Explore next-token prediction interactively. | python gamma.py game |
| Comparison and benchmarking | Compare model behavior, throughput, latency, and code generation. | python gamma.py comparison or python gamma.py benchmark |
| Mind Meld | Coordinate and swap among multiple model engines. | python gamma.py mind-meld |
| SAME-R | Compare prompts, data, teachers, training methods, routers, kernels, or other approaches under matched rims. | projects/samer/README.md |
| Domain research | Run translation, embedding, WGSL, compression, and other owned experiments. | projects/ |
| Integrations | Connect Gamma through APIs, MCP, and supported model engines. | docs/integration-guide.md |
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python gamma.py gameEngine-specific setup is in src/engines/README.md.
SAME-R is Gamma's implementation-neutral outer method: Swappable Approaches under Matched Evaluation and Replication. Its recursive expansion is SAMER Applies Methods, Evaluates, Repeats.
Different inners. Same rims.
The inner approach can be human or agent labeling, data construction, prompt search, SFT, distillation, preference or policy optimization, routing, kernel search, or a future method. SAME-R keeps the objective, evaluator, controls, replication policy, receipts, and promotion boundary matched. A SAME-R instance may itself run as an inner approach.
Gamma currently executes SAME-R through domain project scripts, manifests, and human-selected interventions. The shared automatic approach registry and selector remain an explicit implementation boundary.
- SAME-R project and capability-transfer protocol
- Cross-repository experiment register
- Verifier-guided optimizer and reward taxonomy
The root is limited to stable entrypoints, project metadata, and tool configuration.
| Path | Ownership |
|---|---|
src/ |
Gamma runtime, engines, game, benchmarks, and integrations |
projects/ |
Self-contained research and distillation work |
tools/ |
Operator-facing utilities and model analysis |
requirements/ |
Engine and hardware-specific dependency manifests |
docs/ |
Canonical architecture and usage documentation |
tests/ |
Automated test suites |
Keep generated reports under reports/, run output under its owning project, and
new project-specific utilities beside that project rather than at repository root.
python gamma.py [command]| Command | Purpose |
|---|---|
game |
Interactive gameplay and chat-style runs |
comparison |
Side-by-side model comparison |
mind-meld |
Multi-model collaboration and swapping |
benchmark |
Throughput/latency benchmarking |
codegen |
Code-generation benchmark workflows |
list |
List available models |
select |
Interactive model/engine selector |
help |
Command help |
python gamma.py help
python gamma.py help benchmark
python gamma.py help codegen
python gamma.py game --comparison --help
python gamma.py mind-meld --help- docs/README.md: documentation map and ownership
- docs/BENCHMARKING.md: speed, quality, codegen, and mind-meld benchmarking
- projects/samer/README.md: SAME-R algorithm, controlled capability-transfer trials, and promotion contract
- docs/VERIFIER_GUIDED_LEARNING.md: prompt optimization, verifier-filtered data, RLVR, domain reward boundaries, and experiment states
- projects/samer/experiments/README.md: cross-repository experiment register contract
- docs/integration-guide.md: OpenAI/LangChain/API/MCP integrations
- src/mind_meld/README.md: Mind Meld usage, configs, and operational guardrails
- flux/README.md: Flux install, CLI, and reference
- tools/README.md: project tooling and feedback loop commands
- requirements/README.md: dependency profiles by engine and hardware
- projects/distillation/embedding/README.md: EmbeddingGemma distillation
- projects/distillation/translation/README.md: TranslateGemma distillation
- projects/distillation/wgsl/README.md: Doppler WGSL SFT, DPO, rollout, and GRPO optimizer backend
- src/functiongemma_training/README.md: FunctionGemma training path
MIT. See LICENSE.