Instructions to use polonuim210/ochre-refine-0.8b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use polonuim210/ochre-refine-0.8b with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf polonuim210/ochre-refine-0.8b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf polonuim210/ochre-refine-0.8b:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf polonuim210/ochre-refine-0.8b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf polonuim210/ochre-refine-0.8b:Q4_K_M

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 polonuim210/ochre-refine-0.8b:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf polonuim210/ochre-refine-0.8b:Q4_K_M

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 polonuim210/ochre-refine-0.8b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf polonuim210/ochre-refine-0.8b:Q4_K_M

Use Docker

docker model run hf.co/polonuim210/ochre-refine-0.8b:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use polonuim210/ochre-refine-0.8b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "polonuim210/ochre-refine-0.8b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "polonuim210/ochre-refine-0.8b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/polonuim210/ochre-refine-0.8b:Q4_K_M

Ollama
How to use polonuim210/ochre-refine-0.8b with Ollama:
```
ollama run hf.co/polonuim210/ochre-refine-0.8b:Q4_K_M
```
Unsloth Desktop

How to use polonuim210/ochre-refine-0.8b with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf polonuim210/ochre-refine-0.8b:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "polonuim210/ochre-refine-0.8b:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use polonuim210/ochre-refine-0.8b with Docker Model Runner:
```
docker model run hf.co/polonuim210/ochre-refine-0.8b:Q4_K_M
```

Lemonade

How to use polonuim210/ochre-refine-0.8b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull polonuim210/ochre-refine-0.8b:Q4_K_M

Run and chat with the model

lemonade run user.ochre-refine-0.8b-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use polonuim210/ochre-refine-0.8b with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf polonuim210/ochre-refine-0.8b:Q4_K_M

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 polonuim210/ochre-refine-0.8b:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use polonuim210/ochre-refine-0.8b with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf polonuim210/ochre-refine-0.8b:Q4_K_M

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 "polonuim210/ochre-refine-0.8b:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

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

ochre-refine-0.8b

The local cleanup model for Ochre, open-source voice dictation. It turns raw speech-recognition output into the text the speaker meant to type:

applies self-corrections ("Monday, no wait, Tuesday" → "Tuesday")
removes false starts, stutters and meaningless fillers
fixes punctuation and casing, writes numbers, times, emails and URLs in written form
formats lists only when they were dictated, applies spoken commands ("new line", "comma")
fixes a misrecognized word only when the intended word is clear; otherwise keeps it

It never answers or acts on the text: dictated questions and AI prompts come back as cleaned questions and prompts.

The CPU-only pick.

Files

file	quant	size	notes
`ochre-refine-0.8b-v4-Q4_K_M.gguf`	Q4_K_M	529 MB	current (round 4)
`ochre-refine-0.8b-v3-Q4_K_M.gguf`	Q4_K_M	529 MB	previous release, kept

Training (v4)

A LoRA fine-tune of Qwen/Qwen3.5-0.8B (Apache-2.0, revision 2fc06364715b), merged and quantized to Q4_K_M. Round 4 ("v4").

LoRA r=16, alpha=32, dropout=0.05 on every attention / linear-attention / MLP projection; learning rate 2e-4 (cosine), effective batch 16, 1.5 epochs over the real pairs twice plus the distillation pairs once.
Data: 6,143 audited real-recognizer pairs (upweighted x2) + 23,991 text-only distillation pairs. Real pairs: TTS speech in many voices and acoustic conditions, transcribed by NVIDIA Parakeet TDT 0.6B v3; every target reviewed by two independent auditors plus an adjudicator. Distillation pairs: scripted dictations through a speech-recognition noise simulator fitted to Parakeet's errors, with a 10% two-auditor audit sample.
New in v4: targeted data for the round-3 failures (misrecognized words, long dictations, self corrections, numbers, voice commands, trailing off) and the "recoverability" rule: a word is fixed only when the intended word can be inferred from the transcript.

No user data was used.

Results (blind side-by-side, 3 independent LLM judges, majority vote)

773 held-out dictations (551 from the new round-4 eval set, 222 from earlier rounds); model names hidden, inter-judge Fleiss kappa 0.94.

	v3	v4
pass, all 773	55.6%	63.8%
pass, round-4 eval set (551)	48.8%	59.2%
pass, round-3 held-out set (150)	74.0%	74.7%
pass, round-3 new speech (60)	68.3%	73.3%
harmful failures (lower is better)	25.1%	17.3%

Paired difference v4 - v3: +8.2 [+5.2, +11.3] points (95% bootstrap CI).

Prompt format (important)

Use the production prompt from Ochre (crates/ochre-refine/src/prompts.rs): ChatML, with the dictation wrapped in <dictation> tags and the assistant turn pre-seeded with an empty think block (<think>\n\n</think>\n\n). Use temperature 0 and llama-server's raw /completion endpoint, without --jinja.

Downloads last month: 106

GGUF

Model size

0.8B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for polonuim210/ochre-refine-0.8b

Base model

Qwen/Qwen3.5-0.8B-Base

Finetuned

Qwen/Qwen3.5-0.8B

Quantized

(377)

this model