| --- |
| license: cc-by-nc-4.0 |
| base_model: CohereLabs/tiny-aya-global |
| library_name: peft |
| pipeline_tag: text-generation |
| language: |
| - en |
| tags: |
| - text-to-sql |
| - agent |
| - tool-use |
| - lora |
| - qlora |
| - agenttune |
| - auditkit |
| datasets: |
| - satish9177/tiny-aya-sql-agent-data |
| - trl-lab/SQaLe-2-text-to-SQL-Queries |
| --- |
| |
| # Tiny Aya SQL agent (A1500, supervised fine-tuning only) |
|
|
| A LoRA adapter for Cohere's **Tiny Aya Global**, trained with QLoRA on 1,500 verified multi-turn SQL |
| agent episodes through Lexsi **AgentTune**. |
| - **Its role:** the plain fine-tuning model, the starting point for the submitted model |
| **[B](https://huggingface.co/satish9177/tiny-aya-sql-agent)**. B adds one round of self-training. |
| - **Which to use:** use B for the submission's results, and this one to compare plain fine-tuning with B. |
|
|
| - Write-up: [Teaching Tiny Aya to query databases (HF Article)](https://huggingface.co/blog/satish9177/teaching-tiny-aya-to-query-databases-what-fine-tun) |
| - Code, plans and full results: [github.com/satish9177/aya-db-agent](https://github.com/satish9177/aya-db-agent) |
| - Data, trajectories and logs: [satish9177/tiny-aya-sql-agent-data](https://huggingface.co/datasets/satish9177/tiny-aya-sql-agent-data) |
|
|
| ## Results (strict) |
|
|
| | Set | Base (untuned) | **A1500 (this model)** | B | |
| |---|---:|---:|---:| |
| | SQaLe dev (95) | 2 | **37** | 40 | |
| | SQaLe final (144, unseen databases) | 2.8% | **31.3%** | 30.6% | |
| | BIRD Mini-Dev (493) | 0.0% | **1.6%** | 1.4% | |
| | Shop retention, dev (40) | 0 | **40** | 40 | |
| | Shop retention, multi-hop (40) | 0 | **1** | 0 | |
|
|
| - **On the final sets:** A1500 and B are the same within noise. |
| - **Behaviour:** the same strengths and limits as B. It is a clean tool user for one- and two-table |
| questions. It is weak on 3+ tables, multi-step questions and large schemas, and it never recovers |
| after a failed query. |
|
|
| ## Training |
|
|
| | | | |
| |---|---| |
| | Data | 1,500 verified episodes: 1,265 SQaLe questions + 235 shop-database episodes (1,425 train / 75 validation) | |
| | How the episodes were made | the reference query path replayed against the real database; no teacher model | |
| | Quantization | 4-bit NF4, double quantization, fp16 compute | |
| | LoRA | rank 16, alpha 32, dropout 0.05; q, k, v, o, gate, up and down projections | |
| | Optimization | 2 epochs, learning rate 2e-4, effective batch 16, 180 updates; inputs up to 4,096 tokens | |
| | Loss | the model's own turns only; tool outputs are never trained | |
| | Hardware and cost | one Tesla T4; training plus dev check about $2.36 by rate | |
|
|
| Training records are in `training/`. |
|
|
| ## How to load |
|
|
| Use the same code as on the [B model card](https://huggingface.co/satish9177/tiny-aya-sql-agent), with |
| `"satish9177/tiny-aya-sql-agent-a1500"` as the adapter. The base model is pinned at revision |
| `af89d219b53ed9b13b8a4645f9c8028973510324`. Load the tokenizer from its `tokenizer.json`, as that card |
| explains. |
|
|
| ## Licence and credits |
|
|
| - **Licence:** CC-BY-NC-4.0, following **Tiny Aya Global** by Cohere Labs. Non-commercial use only. |
| - **AgentTune and AuditKIT:** Lexsi Labs, source-available and non-commercial. |
| - **SQaLe:** Wolff, Gomm and Hulsebos (2026). Its card declares MIT. |
|
|
