Instructions to use Kris0404/social-story-gemma4-e2b-v3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Kris0404/social-story-gemma4-e2b-v3 with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("Kris0404/social-story-gemma4-e2b-v3")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use Kris0404/social-story-gemma4-e2b-v3 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Kris0404/social-story-gemma4-e2b-v3"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Kris0404/social-story-gemma4-e2b-v3"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use Kris0404/social-story-gemma4-e2b-v3 with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "Kris0404/social-story-gemma4-e2b-v3"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "Kris0404/social-story-gemma4-e2b-v3"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "Kris0404/social-story-gemma4-e2b-v3",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use Kris0404/social-story-gemma4-e2b-v3 with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Kris0404/social-story-gemma4-e2b-v3"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Kris0404/social-story-gemma4-e2b-v3

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Kris0404/social-story-gemma4-e2b-v3 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Kris0404/social-story-gemma4-e2b-v3"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Kris0404/social-story-gemma4-e2b-v3" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

social-story-gemma4-e2b-v3

LoRA fine-tune of mlx-community/Gemma4-E2B-IT-Text-int4 (text-only Gemma 4 E2B Instruct, 4-bit) trained to generate Carol-Gray-style social stories for autistic children. Fused, then quantized for both Python (mlx-lm) and Swift (mlx-swift-lm) on-device deployment.

Use case

Drop-in on-device text generation backend for the Mystoria iOS/iPadOS app (formerly Social Story Studio). Pairs with the v2-era prompt template — same system message and 8-rule user prompt — and emits valid JSON {title, pages: [{text, order}]}.

Training

Base: mlx-community/Gemma4-E2B-IT-Text-int4 (Gemma 4 E2B Instruct, text-only, 4-bit)
Method: LoRA via mlx_lm.lora
Hyperparameters: rank 8, scale 16, num_layers 26, lr 1e-5, batch 1 + grad-accum 4, max_seq 1024
Iters: 2500
Dataset: ~1,720 train / 152 valid / 153 test social stories generated and filtered to first-person, qualifier-rich, tentative-directive Carol Gray voice (10.2 / 10.4 methodology)

Eval (20-prompt grid: 5 ages × 4 topics)

Metric	Rate
JSON valid	100%
Schema correct (text/order)	100%
Tier-1 Carol Gray pass	70%
Descriptive:coaching ≥ 2:1	80%
Test loss	0.447 (ppl 1.563)

The 30% Tier-1 misses are all driven by the descriptive:coaching ratio on coaching-heavy topics (e.g. "making a new friend"). Every other sub-check (first-person, forbidden phrases, qualifiers, no-figurative-language, structure, positive framing, reading level, no-page-dupes) passes 100%.

Quantization (mixed precision for Swift compatibility)

Most weights are 4-bit (group_size=64, ~4.5 bits/weight). The per_layer_model_projection layer is kept in float16 because mlx-swift-lm 3.31.3's Gemma4Text.ScaledLinear is a plain Module and cannot dequantize at load time (the loader sees the packed (8960, 192) uint32 shape and rejects). Keeping that one layer in fp16 adds ~25 MB but lets the model load on iOS without a Swift-side patch.

Deployment

Final size: ~2.5 GB on disk
Loads via standard mlx_lm.load (Python) and mlx-swift-lm ≥ 3.31.3 (Swift) — both work without the sanitize patch the original mlx-community/Gemma4-*-Text-int4 checkpoints needed
iOS users: clear any prior cache from a broken v3 download before loading this revision

Limitations

The descriptive:coaching ratio fails on topics where every "I can try to..." sentence reads as coaching. Future work: more descriptive-first training data for action-oriented topics.
On-device generation is text-only. Image generation in the app continues to run cloud-only.

Downloads last month: 22

Safetensors

Model size

5B params

Tensor type

U32

F16

MLX

Hardware compatibility

4-bit

Model tree for Kris0404/social-story-gemma4-e2b-v3

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Adapter

(183)

this model