Downloads · 30 days
0
pivotai-curriculum-lora
ishreyadev/pivotai-curriculum-lora
pivotai-curriculum-lora is a machine learning model from ishreyadev. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
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 (compl…
Access
Public
Updated Jul 10, 2026
Repo size
101 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors83.9 MB · 83%
At a glance
- License
- apache-2.0
- Access
- Public
- Created
- Jul 10, 2026
- Updated
- Jul 10, 2026
- SHA
- 4a790ae0
Base models
Model card
From the Hugging Face model README
pivotai-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 pivotai 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 pivotai-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 pivotai-curriculum -f Modelfile.curriculum
ollama run pivotai-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.
- License
- apache-2.0
- Languages
- en
- Created
- Jul 10, 2026
- Updated
- Jul 10, 2026