pivotai-distill-lora

PivotAI-distill

Knowledge-distilled Llama 3.1 8B for Indian domestic travel optimization. Distilled from 500 multi-agent DeepSeek reasoning traces (Phase 2 of the PivotAI pipeline), where a Supervisor + Analyst + Concierge + Optimizer chain used real MCP tool calls to build itineraries.

Part of the PivotAI project. Unlike PivotAI-ft (trained on clean synthetic pairs), this model was trained on agent reasoning chains — the hypothesis being that richer teacher signal improves generalization. Results were mixed: reasoning coherence improved, but structural output compliance dropped.

Model Details

The higher loss (0.429 vs 0.266 for ft) correlates with noisier training signal — agent traces include tool-call artifacts and variable output lengths that add training noise.

Evaluation Results (92 test cases)

Schema compliance of 0% indicates the model produces valid JSON but with a different structure than the expected schema — a consequence of the diverse output formats in the distillation training data.

Usage with Ollama

ollama create PivotAI-distill -f Modelfile.distill
ollama run PivotAI-distill

Prompt format (Alpaca with reasoning chain instruction):

### Instruction:
Act as PivotAI Supervisor for an Indian domestic trip. Coordinate the Analyst, Concierge, and Optimizer agents to find Price-Pivot Points and produce an optimized itinerary. Show the reasoning chain for each agent handoff, then provide the final pivot analysis and optimized itinerary.

### Input:
{"starting_city": "Mumbai", ...}

### Response:

Limitations

• Schema compliance is 0% — produces valid JSON but in a non-standard structure.
• Not recommended for production use without post-processing to extract the itinerary.
• Trained on only 449 examples (vs 4,749 for ft) — limited coverage of edge cases.