Advanced LLM Techniques: Fine-Tuning for OpenClaw

Clawpedia · For Humans

Fine-tune language models specifically for OpenClaw to improve performance on your custom tasks.

Specialized Models for Specialized Tasks

While general-purpose models like GPT-4o and Claude work well for most tasks, fine-tuning creates a model that excels at your specific use case. This guide explains when and how to create fine-tuned models that integrate with OpenClaw.

---

When to Fine-Tune

Cost-Benefit Analysis

---

Preparing Training Data

Format

Training data uses the JSONL format with conversation examples:

{"messages": [{"role": "system", "content": "You are a medical coding assistant."}, {"role": "user", "content": "Patient has type 2 diabetes with peripheral neuropathy"}, {"role": "assistant", "content": "ICD-10 Codes:\n- E11.40: Type 2 diabetes mellitus with diabetic neuropathy, unspecified\n- G63: Polyneuropathy in diseases classified elsewhere"}]}
{"messages": [{"role": "system", "content": "You are a medical coding assistant."}, {"role": "user", "content": "Patient presents with acute bronchitis"}, {"role": "assistant", "content": "ICD-10 Code:\n- J20.9: Acute bronchitis, unspecified"}]}

Data Requirements

Generating Training Data from OpenClaw

Use your best conversations as training examples:

# Export conversations that produced good results
openclaw memory export --category conversations --format jsonl > training_data.jsonl

# Filter for high-quality examples
openclaw say "Filter this JSONL file to only include conversations where the user gave positive feedback" < training_data.jsonl > filtered_data.jsonl

---

Fine-Tuning with OpenAI

Step 1: Upload Training Data

curl https://api.openai.com/v1/files \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -F purpose="fine-tune" \
  -F file="@training_data.jsonl"

Step 2: Start Fine-Tuning

curl https://api.openai.com/v1/fine_tuning/jobs \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "training_file": "file-abc123",
    "model": "gpt-4o-mini-2024-07-18",
    "suffix": "medical-coding"
  }'

Step 3: Monitor Training

curl https://api.openai.com/v1/fine_tuning/jobs/ftjob-xxx \
  -H "Authorization: Bearer $OPENAI_API_KEY"

Step 4: Use in OpenClaw

openclaw config set model ft:gpt-4o-mini-2024-07-18:org:medical-coding:abc123
openclaw restart

---

Evaluation

Always evaluate your fine-tuned model before deployment:

# Create a test set (separate from training data)
# Run both base and fine-tuned models on the same inputs
# Compare accuracy, consistency, and quality

openclaw say "Test query" --model gpt-4o-mini > base_output.txt
openclaw say "Test query" --model ft:gpt-4o-mini:...:abc123 > finetuned_output.txt

diff base_output.txt finetuned_output.txt

Metrics to Track

---

Tips

• Start with prompt engineering before fine-tuning — it is faster and cheaper.
• Use at least 50 high-quality examples for meaningful improvement.
• Separate training and test data to avoid overfitting.
• Fine-tune the smallest model that meets your needs — it is cheaper and faster.
• Version your training data alongside your model versions.
• Re-fine-tune periodically as your requirements evolve.

---

Troubleshooting