Instructions to use 4rc4n4/qwen-3-4b-guru with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 4rc4n4/qwen-3-4b-guru with Transformers:

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

pipe = pipeline("text-generation", model="4rc4n4/qwen-3-4b-guru")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("4rc4n4/qwen-3-4b-guru", dtype="auto")

llama-cpp-python

How to use 4rc4n4/qwen-3-4b-guru with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="4rc4n4/qwen-3-4b-guru",
	filename="gguf/qwen-3-4b-guru-F16.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use 4rc4n4/qwen-3-4b-guru with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf 4rc4n4/qwen-3-4b-guru:F16
# Run inference directly in the terminal:
llama-cli -hf 4rc4n4/qwen-3-4b-guru:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf 4rc4n4/qwen-3-4b-guru:F16
# Run inference directly in the terminal:
llama-cli -hf 4rc4n4/qwen-3-4b-guru:F16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf 4rc4n4/qwen-3-4b-guru:F16
# Run inference directly in the terminal:
./llama-cli -hf 4rc4n4/qwen-3-4b-guru:F16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf 4rc4n4/qwen-3-4b-guru:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf 4rc4n4/qwen-3-4b-guru:F16

Use Docker

docker model run hf.co/4rc4n4/qwen-3-4b-guru:F16

LM Studio
Jan

vLLM

How to use 4rc4n4/qwen-3-4b-guru with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "4rc4n4/qwen-3-4b-guru"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "4rc4n4/qwen-3-4b-guru",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/4rc4n4/qwen-3-4b-guru:F16

SGLang

How to use 4rc4n4/qwen-3-4b-guru 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 "4rc4n4/qwen-3-4b-guru" \
    --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": "4rc4n4/qwen-3-4b-guru",
		"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 "4rc4n4/qwen-3-4b-guru" \
        --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": "4rc4n4/qwen-3-4b-guru",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use 4rc4n4/qwen-3-4b-guru with Ollama:
```
ollama run hf.co/4rc4n4/qwen-3-4b-guru:F16
```

Unsloth Studio

How to use 4rc4n4/qwen-3-4b-guru with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for 4rc4n4/qwen-3-4b-guru to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for 4rc4n4/qwen-3-4b-guru to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for 4rc4n4/qwen-3-4b-guru to start chatting

How to use 4rc4n4/qwen-3-4b-guru with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf 4rc4n4/qwen-3-4b-guru:F16

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "4rc4n4/qwen-3-4b-guru:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use 4rc4n4/qwen-3-4b-guru with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf 4rc4n4/qwen-3-4b-guru:F16

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 4rc4n4/qwen-3-4b-guru:F16

Run Hermes

hermes

Docker Model Runner
How to use 4rc4n4/qwen-3-4b-guru with Docker Model Runner:
```
docker model run hf.co/4rc4n4/qwen-3-4b-guru:F16
```

Lemonade

How to use 4rc4n4/qwen-3-4b-guru with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull 4rc4n4/qwen-3-4b-guru:F16

Run and chat with the model

lemonade run user.qwen-3-4b-guru-F16

List all available models

lemonade list

qwen-3-4b-guru

A fast chunk→concept tagger for the guru comparative-religion indexing pipeline. Fine-tuned from Qwen3-4B-Instruct-2507 on 2,808 (passage, tag-set) pairs — teacher labels from a 27B model, refined by human review — this model scores passages from mystical texts against a curated 88-concept taxonomy.

Exploratory v1. This release was built to test Qwen3-4B as a faster replacement for the bulk-tagging step. It met no formal ship gates (none were set); see Evaluation and Limitations for exactly what was and wasn't measured.

What it does

Given a passage and a list of candidate concepts (each {id, definition}), the model returns a JSON array rating every present concept 0–3 (0=absent … 3=central theme); concepts scoring 0 are omitted. Output is strict JSON, no prose. The prompt contract matches the guru tagging caller exactly.

Evaluation

Held-out test split (293 chunks the model never trained on), scored against both the 27B teacher's labels and independent human accept/reject verdicts. base = Qwen3-4B-Instruct-2507 with no fine-tuning. Both at temperature 0 via llama-server, identical prompts.

vs human-graded labels (1,990 graded cells)

Strongest signal — human verdicts are an independent ground truth.

Model	Precision	Recall	F1	Specificity
base	0.682	0.550	0.609	0.590
qwen-3-4b-guru	0.710	0.571	0.633	0.626
27B teacher (reference)	0.616	1.000*	0.762*	0.000

* The teacher's recall is 1.0 by construction — humans only reviewed tags the teacher emitted — so its precision (0.616) is the comparable number. This model's precision (0.710) clears it.

vs teacher labels

Model	Precision	Recall	F1	Macro-F1	Parse rate	OOT-IDs	tags/chunk	Lat (s)
base	0.345	0.561	0.427	0.293	94.9%	18	27.6	13.9
qwen-3-4b-guru	0.557	0.614	0.584	0.432	99.7%	1	12.5	5.3

The fine-tune's main effect is calibration, reliability, and speed: the base model massively over-tags (27.6 concepts/chunk, 18 invented IDs, 94.9% parse, 13.9 s) while this model is disciplined (12.5 concepts/chunk, 1 invented ID, 99.7% parse, 5.3 s — ~2.6× faster). It modestly improves on the base against human truth too (F1 0.609 → 0.633).

Training

Base: unsloth/Qwen3-4B-Instruct-2507-bnb-4bit
Method: QLoRA (TRL SFTTrainer via Unsloth)
LoRA: r=32, α=64, dropout 0, all attention + MLP projections
Schedule: 3 epochs, batch 1 × grad-accum 16, paged AdamW-8bit, lr 1.5e-4 cosine, warmup 0.03
Sequence length: 6144 (8192 OOM'd on a 24 GB 3090 — the loss logits are seq_len × 151,936 vocab; 6144 fits and drops only ~5% of examples)
Chat template: Qwen3 ChatML (qwen3-instruct)
Checkpoint: best-by-val-loss (eval_loss 0.3421)
Hardware / wall-clock: single RTX 3090 / ~18 h
Seed: 42

Training data

Source: staged_tags from a guru.db snapshot, teacher Qwen3.5-27B-UD-Q4_K_XL.gguf, prompt version v1, status ∈ {pending, accepted} (human-rejected tags dropped).

2,808 chunks, 88 concepts, 25,893 target tags (47% human-accepted, rest unreviewed teacher labels, 0 rejected)
Splits 80/10/10 by chunk_id (2,245 / 270 / 293); 111 train chunks (4.9%) dropped for exceeding 6144 tokens

The 88-concept taxonomy this model expects is pinned in taxonomy.toml in this repo. (The upstream guru taxonomy has since migrated to a three-tier hierarchy; this model's contract is the flat 88-concept version it trained on.)

Files

adapter/ — LoRA adapter (~260 MB); the canonical artifact
merged/ — adapter merged into base, FP16 (~8 GB)
gguf/qwen-3-4b-guru-Q4_K_M.gguf — 4-bit, ~2.5 GB, recommended for serving
gguf/qwen-3-4b-guru-F16.gguf — full-precision conversion
taxonomy.toml — the 88-concept taxonomy (prompt contract)

Usage

llama-server -m qwen-3-4b-guru-Q4_K_M.gguf --jinja --port 8080

The guru tagging caller hits the OpenAI-compatible /v1/chat/completions endpoint. The model expects the exact prompt structure used at training time (system role + passage + 0–3 rubric + JSON concept list); deviating degrades quality.

Limitations

Exploratory release. No formal throughput benchmark and no quantization sweep were run; only Q4_K_M is provided and serving throughput at concurrency is unmeasured. Treat the eval as a sound point estimate, not a gated guarantee.
Domain-locked. The corpus is heavily Mediterranean / Greek-philosophical; calibration on East-Asian, South-Asian, and indigenous traditions is weaker.
Taxonomy-bound. Scoring is conditioned on the 88-concept list in the prompt. Use the pinned taxonomy.toml; if you change the taxonomy meaningfully, retrain.
Label noise. ~53% of training targets are unreviewed teacher labels, so the model inherits some of the teacher's tagging tendencies on un-reviewed concepts.
Not a chat model. Tuned on a single task and prompt format.

License

Apache 2.0

Downloads last month: 34

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

4-bit

16-bit

Model tree for 4rc4n4/qwen-3-4b-guru

Base model

Qwen/Qwen3-4B-Instruct-2507

Quantized

unsloth/Qwen3-4B-Instruct-2507-bnb-4bit

Adapter

(5)

this model