gemma-e2b-rlcd

Gemma E2B RLCD

Parallel classification, grading, and label probabilities over text, images, speech, and video, powered by Gemma 4 E2B on Apple Silicon.

Give the model one input and a set of questions. It encodes the input once, scores the allowed answers in GPU batches, and returns structured results with probability distributions.

This download includes the complete 4-bit multimodal checkpoint and the gemma_rlcd runtime. The checkpoint preserves the pinned MLX quantization; the runtime implements parallel decision scoring. Source code on GitHub.

<video controls playsinline preload="metadata" width="100%" src="https://huggingface.co/larkooo/gemma-e2b-rlcd/resolve/main/docs/assets/live-demo.mp4"></video>

128 visual checks, streaming live. In this recorded simultaneous run: 11.57 s vs 54.10 s · 4.67× faster · 120/128 matching answers. Both paths share one GPU. The full startup pause and original elapsed time are preserved.

Download the video · Recorded results

Quick start

Requires an Apple Silicon Mac, Python 3.12+, uv, and ffmpeg for audio/video.

uvx --from huggingface_hub hf download larkooo/gemma-e2b-rlcd \
  --local-dir gemma-e2b-rlcd
cd gemma-e2b-rlcd
uv venv --python 3.12
source .venv/bin/activate
uv pip install '.[web]'
brew install ffmpeg

gemma-rlcd-web --model . --port 8787

Open http://127.0.0.1:8787. Add text, upload media, or record speech. Enter a question and its choices—“Which brand is this?” with Porsche, Mercedes, and Other is enough. Descriptions are optional.

Choose Run all fields for answers and distributions, or Compare with Gemma for a side-by-side comparison with ordinary JSON generation. Built-in examples cover support triage, customer inboxes, incident facts, and catalog routing. Uploads stay local and are deleted after each request.

The download is approximately 3.6 GB and includes the image and audio encoders, tokenizer, processor, and chat template. Inference runs locally on Apple Silicon through MLX.

Live visual demo

Open http://127.0.0.1:8787/demo for 32, 64, or 128 checks over an image or video. Use the included street photo or upload your own media. The parallel scorer streams completed field batches; normal Gemma streams its generated JSON. Live clocks, per-check probabilities, answer differences, and downloadable events make the comparison inspectable. Both paths start together and stream side by side, sharing the resident weights with separate processors and KV caches. Timings measure concurrent completion on one GPU, including resource contention. The playground’s ordinary comparison remains sequential for isolated timings.

The first answer follows input preparation and a full multimodal prefill. The UI shows these stages, first-token time, and first-decision time separately. Repeated runs reuse compiled field definitions; media, input KV state, and answers are recomputed. Use Focus view to see all 128 outputs together.

How it works

flowchart LR
    A[Text / image / speech / video] --> B[Shared multimodal prefill]
    Q[Questions and allowed answers] --> B
    B --> C[Reusable KV state]
    C --> D[Batched field and candidate scoring]
    D --> E[Probabilities and structured results]

1. Encode once. Gemma processes the complete input, questions, and candidate descriptions in one shared prefix.
2. Branch by field. Each field starts from that prefix with its own short JSON key. Branches run in bounded GPU batches.
3. Score complete answers. Single-token choices share one output vector. Multi-token choices are evaluated as complete JSON values with teacher forcing, including closing quotes.
4. Return the result. Candidate log likelihoods become probabilities within each field. Code selects labels, calculates grades, and assembles the final JSON.

The output values are known in advance, so inference can score them without an autoregressive answer-generation loop. All 35 Gemma layers and the native image/audio processing remain in use. Video includes sampled frames, timestamps, and its soundtrack.

The default batch size is eight. More fields use additional batches; prefix computation is reused, while KV storage is replicated per batch row. Fields share the complete input and schema, but do not condition on one another's generated answers. See architecture for the scoring equations, precision settings, and execution details.

RLCD and probability learning

RLCD stands for Reinforcement Learning for Calibrated Decisions: learning decision probabilities from outcome feedback. A calibrated 80% prediction should succeed about 80% of the time across comparable cases.

Here, the default inference path obtains candidate probabilities directly from pretrained Gemma. The repository also includes a candidate-conditioned decision head, supervised likelihood training, Brier/log-loss evaluation, and temperature fitting. Training and calibration explains these components and the outcome-feedback objective separately from inference.

Output types

For independent labels, {cat: 0.9, dog: 0.5} is valid: both can be present. A mutually exclusive cat-or-dog choice sums to 1. Scores use zero-based grade levels; confidence measures distribution concentration via normalized entropy. Probability quality can be evaluated with the calibration utilities.

from gemma_rlcd import Choice, DecisionEngine, Independent, State
from gemma_rlcd.json_backend import JSONMLXBackend

engine = DecisionEngine(JSONMLXBackend("."))
result = engine.system_one(
    State(text="A cat sleeps on the sofa. No dogs are present."),
    {
        "animal": Choice("Which animal is present?", {"cat": "cat", "dog": "dog"}),
        "presence": Independent(
            "Which animals are present?",
            {"cat": "A cat is present", "dog": "A dog is present"},
        ),
    },
)
print(result["answers"])

State also accepts images, audio, and videos as tuples of local paths. Run a JSON request from the command line with:

gemma-rlcd examples/text.json --model .

Performance

Local Apple M5 measurements: the same 4-bit model, four measured runs after warmup, fresh KV state per request. Times include input preparation and inference; model loading and upload are excluded.

A ratio above 1 favors parallel scoring. The support example matches all 28 outputs; 17 have annotated expectations. Many short decisions benefit most, while long candidate sets add scoring work. The full comparison includes all workloads, timing ranges, answer differences, and a serial scoring control. The demo replays these measured support-case medians.

Media and request limits

The playground accepts up to eight images, one audio source, and one video, with 200 MB of uploads per request. Audio and videos with sound support up to 30 seconds; silent video supports up to 60 seconds. Video targets one frame per second, capped at 32 frames, so brief events can fall between samples.

Requests support up to 32 named fields, 128 primitive decisions, and 8,192 processed input tokens. Oversized inputs return an error rather than being truncated. Multi-picture phone JPEGs use the full-resolution primary photograph. Text and field definitions are saved in browser local storage; uploaded media is not retained.

Development

python -m pytest -q
ruff check gemma_rlcd tests scripts
ruff format --check gemma_rlcd tests scripts
node --check gemma_rlcd/static/app.js

See contributing, training, and experiment reports. The tested dependencies and model revision make the local setup reproducible. CI runs portable tests on Linux; MLX inference and GPU tests run on Apple Silicon.

License

Gemma 4 weights: Apache 2.0, as specified by Google. Runtime code and original synthetic examples: MIT. The checkpoint manifest records source revisions and file hashes. Third-party code and adapted examples are documented in notices.