djx-local-model

name	djx_local_model
description	Data-Juicer local model processing: Ollama configuration, private data handling, offline workflows. Trigger keywords: local model, Ollama, private data, sensitive data, offline, do not send to cloud, local LLM. Use when processing sensitive/private data, running offline, or configuring local models. Related skills: djx_auth (authentication), data-juicer (main flow), djx_retrieve (local retrieval).
allowed-tools	Bash, Read
argument-hint
user-invocable	true

Data-Juicer Skills: Local Model Processing

Process sensitive or private data — all operations use local models; data never leaves the machine.

---

Core Rule: Use djx tool Only

You must only use the djx tool CLI. Do not use the session or cap modules.

---

Prerequisites

Condition	Requirement	Verification Command
Ollama installed	Installed and running	ollama --version
Model pulled	At least one model downloaded	ollama list
Service running	Ollama service is running	curl http://localhost:11434/v1/models

---

Environment Configuration

Route all LLM calls to local Ollama:

export DJA_OPENAI_BASE_URL="http://localhost:11434/v1"
export DASHSCOPE_API_KEY="ollama"
export DJA_SESSION_MODEL="qwen3.5:0.8b"
export DJA_PLANNER_MODEL="qwen3.5:0.8b"
export DJA_LLM_THINKING="false"

Variable	Reason for Value
DJA_OPENAI_BASE_URL	Ollama's OpenAI-compatible endpoint
DASHSCOPE_API_KEY	Placeholder — Ollama does not validate keys
DJA_SESSION_MODEL / DJA_PLANNER_MODEL	Must match a pulled Ollama model
DJA_LLM_THINKING	false — local models do not support extended thinking

---

Ollama Setup

Start Service

ollama serve &

Pull Models

# Lightweight — fast, low resources
ollama pull qwen3.5:0.8b

# Balanced — better quality
ollama pull qwen3.5:2b

# High quality — best results, more resources
ollama pull qwen3.5:4b

Verify

# List pulled models
ollama list

# Test endpoint
curl http://localhost:11434/v1/models

---

Model Selection Guide

Hardware	Recommended Model	Memory Usage
8GB RAM	qwen3.5:0.8b	~1GB
16GB RAM	qwen3.5:2b	~2GB
32GB+ RAM	qwen3.5:4b	~4GB

---

Local Processing Guide

Fully Local Tools (No LLM Required)

These tools run entirely locally:

Tool	Function
inspect_dataset	Dataset analysis
build_process_spec	Spec validation
build_dataset_spec	Spec building
apply_recipe	Execution via dj-process

Local Operator Retrieval

retrieve_operators runs entirely locally — no API key, no cloud calls:

djx tool run retrieve_operators --input-json '{"intent": "<goal>", "top_k": 5}'

After selecting a candidate, use get_operator_info to inspect its parameter schema:

djx tool run get_operator_info --input-json '{"operator_name": "<operator_name>"}'

Semantic Operators Using Local Models

When the pipeline includes LLM-based operators, configure them to use the local endpoint.

Operator parameters come from retrieve_operators output. If parameters are unclear, use empty {} — the operator will apply default values.

---

Error Handling

Scenario	Solution
Connection refused	Start Ollama: ollama serve &
model not found	Pull the model: ollama pull <model>
Port conflict (11434 in use)	Use a different port: OLLAMA_HOST=0.0.0.0:11435 ollama serve &, then update DJA_OPENAI_BASE_URL
DJA_LLM_THINKING=true error	Set to false for all local models
auto mode 401 fallback	Use retrieve_operators (already local, no API needed)

---

Must-Read Pitfalls

1. Always Verify Ollama Is Running

ollama list
curl http://localhost:11434/v1/models

2. retrieve_operators Is Already Local

No need to specify a special mode — retrieve_operators runs entirely locally by default:

djx tool run retrieve_operators --input-json '{"intent": "...", "mode": "auto"}'

3. Never Send Private Data to the Cloud

Configure all semantic operators to use the local endpoint.

4. Test with Small Samples Before Processing

Validate the pipeline with a small sample before processing the full dataset.

5. Pull Models in Advance

Pull models before starting the workflow to avoid mid-process downloads.

---

Complete Local Workflow Example

# 1. Start Ollama
ollama serve &

# 2. Pull model (first time)
ollama pull qwen3.5:0.8b

# 3. Configure environment
export DJA_OPENAI_BASE_URL="http://localhost:11434/v1"
export DASHSCOPE_API_KEY="ollama"
export DJA_SESSION_MODEL="qwen3.5:0.8b"
export DJA_PLANNER_MODEL="qwen3.5:0.8b"
export DJA_LLM_THINKING="false"

# 4. Verify
ollama list
curl http://localhost:11434/v1/models

# 5. Inspect dataset
djx tool run inspect_dataset --input-json '{"dataset_source": {"path": "/data/sensitive.jsonl"}, "sample_size": 50}'

# 6. Retrieve operators locally (already local, no API key needed)
djx tool run retrieve_operators --input-json '{"intent": "clean and filter sensitive data", "top_k": 10}'

# 7. Continue with main flow...

---

Skill Responsibilities

Scenario	Skill to Use
Configure API Key (cloud)	djx_auth
Configure local models	djx_local_model (this skill)
Local operator retrieval	djx_retrieve (retrieve_operators, already local)
Main flow	data-juicer