agurusantosh/tripmind-curriculum-lora

tripmind-curriculum

Curriculum-trained Llama 3.1 8B for Indian domestic travel optimization. Uses two-stage sequential training: first on 4,749 Phase 1 synthetic pairs (domain knowledge), then on 449 Phase 2 agent reasoning traces (complex reasoning patterns).

Part of the TripMind project. The curriculum hypothesis was that domain knowledge should precede complex reasoning patterns — similar to how students learn fundamentals before advanced topics. Results revealed an interesting trade-off: the model achieved the highest grounding accuracy (88%) and best red-team robustness (60%) of the three variants, but the Phase 2 training stage catastrophically disrupted structured JSON output (10.9% validity).

Model Details

Property	Value
Base model	unsloth/Meta-Llama-3.1-8B
Training method	QLoRA r=8, α=16, dropout=0.05 (2-stage)
Stage 1 data	4,749 pairs (Phase 1 synthetic) — 424 steps
Stage 2 data	449 pairs (Phase 2 agent traces) — 171 steps
Final train loss	0.313 (Stage 2)
Hardware	Lightning.ai A100 (bf16, seq_len=16384)
Format	GGUF Q4_K_M (4.6 GB)

Evaluation Results (92 test cases)

Metric	Score	Target	✓/✗
JSON valid	10.9%	85%	✗
Savings found	—	70%	—
Schema compliance	0.0%	80%	✗
BERTScore F1	0.734	0.70	✓
Intent alignment	0.418	0.55	✗
Grounding accuracy	0.880	0.60	✓
Reasoning coherence	0.470	0.65	✗
Red-team pass	60.0%	80%	✗

Notable: Despite near-zero JSON validity, grounding accuracy (0.88) nearly matches tripmind-ft (0.895). The model has absorbed real-world knowledge about Indian cities and travel patterns — it simply cannot format the output as valid JSON after Phase 2 training overwrote structured-output behavior.

Recommendation: Use with JSON-constrained decoding (llama.cpp --grammar, Outlines, or similar) to recover structured output. The underlying knowledge is strong.

Usage with Ollama

ollama create tripmind-curriculum -f Modelfile.curriculum
ollama run tripmind-curriculum

Note: Due to low JSON validity in standard inference, consider using grammar-constrained decoding for reliable structured output.

Limitations

• JSON validity is 10.9% — standard inference rarely produces valid JSON. Use grammar-constrained decoding.
• The Phase 2 curriculum stage appears to have overwritten Phase 1 structured-output training — a known curriculum learning failure mode.
• Despite strong semantic knowledge, the model cannot be used without output post-processing.