Instructions to use rockus/Poocha-E4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use rockus/Poocha-E4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="rockus/Poocha-E4B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("rockus/Poocha-E4B")
model = AutoModelForMultimodalLM.from_pretrained("rockus/Poocha-E4B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use rockus/Poocha-E4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "rockus/Poocha-E4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rockus/Poocha-E4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/rockus/Poocha-E4B

SGLang

How to use rockus/Poocha-E4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "rockus/Poocha-E4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rockus/Poocha-E4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "rockus/Poocha-E4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "rockus/Poocha-E4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use rockus/Poocha-E4B with Docker Model Runner:
```
docker model run hf.co/rockus/Poocha-E4B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Poocha-E4B — a kids' science tutor

Poocha (പൂച്ച, "cat") is a family of small, child-friendly science-tutor models. The assistant persona is Poocha, a clever, curious kitten who teaches children (ages ~8-12) about space, oceans, plants and animals using everyday Indian analogies.

This is the E4B variant: a bf16 LoRA fine-tune of google/gemma-4-E4B-it (~4.5B effective (8B w/ embeddings)), merged to full precision.

Training

Base: google/gemma-4-E4B-it (Gemma 4, Apache-2.0)
Method: bf16 LoRA (r=16, α=16, dropout 0.05), 2 epochs, lr 1e-4, on an NVIDIA B200 (Unsloth). Round-3 metrics: train_loss ≈ 0.226, eval_loss ≈ 0.799 (both improved over the 2.9k-row Round-2 set despite 4× the data).

Training corpus (~12.1k chat rows, every row in Poocha's voice)

A fresh, sanitized multi-corpus set — four parts, all {"messages": [...]} chat:

Reused first round (~2.7k) — persona-rich Q&A + Indian-context stories + multi-turn dialogues (the Round-2 set, cleaned and persona-renamed Kat→Poocha).
Factual rewrite (~9.0k) — every encyclopedic passage from NCERT Science 6–9, Science Journal for Kids, and Tushe/Siyavula re-narrated in Poocha's own first-person voice (style transfer, not raw text), with a teacher fact-check pass.
Adventures (~0.4k) — interactive, science-grounded mini-adventures for the UI's "Adventure mode".
Behaviours (~0.08k) — identity, gentle off-topic redirects, kid-safe handling, world curiosity, and image-drawing, seeded from real usage logs.

Data design: facts are always re-told in Poocha's voice (never trained on raw text — the lesson from a failed early round); a deterministic sanitizer drops leaked numbers, source/credit junk, truncations and visual/factual hallucinations; the engagement "what should we explore next?" close is reinforced because real kids reliably follow it. Teacher = Gemma 4 31B FP8 (vLLM).

Evaluation

ARC-Challenge-Indic (English): 90.5% (181/200) — science knowledge.
Engagement loop ("what should we explore next?") present in 92% of sampled answers — the behaviour real kids follow.
0 hallucinations flagged by the deterministic guards and 0% dry (persona always present) on the behaviour suite.

GGUF / local serving

Quantized GGUFs (F16 · Q8_0 · Q6_K) for llama.cpp / Ollama / LM Studio live in rockus/Poocha-E4B-GGUF. Q6_K (6.2 GB) is the recommended deploy and fits a 12 GB GPU with KV headroom.

Usage (recommended sampling)

Factual Q&A: temperature=0.30, min_p=0.08, top_k=0, top_p=1.0
Storytelling: temperature=0.95, min_p=0.05, top_k=0, top_p=1.0
Add repetition_penalty≈1.15 for cleaner long outputs.

System prompt:

You are Poocha, a clever, curious little kitten who teaches Indian children (ages 8-12) about science. Warm, encouraging, plain-spoken. You may use a gentle purr or meow OCCASIONALLY. Use simple Indian examples.

Limitations

Trained on synthetic data (generated/rewritten by Gemma 4 31B); fact-checked by the teacher + a deterministic sanitizer, but not exhaustively verified — occasional minor coherence wobbles remain.
Grounding spans NCERT (Indian, classes 6–9), Science Journal for Kids, and Tushe/Siyavula (South-African open textbooks); some framing reflects those sources.
Kid-targeted (ages ~8–12); English only; persona is intentional and pervasive. E2B is sized for on-device/tablet; E4B is the recommended, higher-quality variant.

License

Apache-2.0 (inherited from Gemma 4).

Downloads last month: -

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for rockus/Poocha-E4B

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Finetuned

(226)

this model

Quantizations

1 model

rockus
/

Poocha-E4B