sql-debug-agent-qwen25-05b-grpo-wandb-continue-v2/README.md

---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- text-generation
- conversational
- qwen2
- trl
- grpo
- safetensors
- text-generation-inference
base_model:
- Qwen/Qwen2.5-Coder-0.5B-Instruct
- Qwen/Qwen2.5-Coder-7B-Instruct
model-index:
- name: sql-debug-agent-qwen25-05b-grpo-wandb-continue-v2
  results:
  - task:
      type: text-generation
      name: SQL Repair (Execution-Grounded)
    dataset:
      type: openenv-sql-debug
      name: SQL Debug Environment task suite
    metrics:
    - type: spider_style_headline
      value: 78.5
      name: Spider-style headline
---

# Model Card for `md896/sql-debug-agent-qwen25-05b-grpo-wandb-continue-v2`

## Model Details

| Field | Value |
|---|---|
| Developed by | Md Ayan (`mdayan8`) |
| Model type | Causal LM fine-tuning workflow for SQL debugging/repair |
| Language | English (SQL + natural language prompts) |
| License | Apache-2.0 |
| Shared by | `md896` |
| Pipeline tag | Text Generation |
| Model family tags | `qwen2`, `trl`, `grpo`, `conversational`, `text-generation-inference` |

## Model Description

This model is part of an execution-grounded SQL debugging workflow built on OpenEnv tasks. The key idea is to optimize for runtime correctness rather than only text-level plausibility.

The training/evaluation workflow uses:

1. A fast bridge phase on **Qwen2.5-Coder-0.5B-Instruct** for environment wiring checks.
2. Baseline/eval track with **Qwen2.5-Coder-7B-Instruct** and benchmark comparisons.
3. GRPO-based optimization signals from SQL execution outcomes, grader feedback, and task completion behavior.

## Model Sources

- Repository: https://github.com/mdayan8/sql-debug-env
- Demo / Environment: https://md896-sql-debug-env.hf.space
- Training dashboard (W&B): https://wandb.ai/mdayanbag-pesitm/sql-debug-grpo-best-budget/workspace?nw=nwusermdayanbag
- Reference arXiv listed for metadata context: https://arxiv.org/abs/1910.09700

## Intended Uses

### Direct Use

- SQL repair assistant style prompting in controlled environments
- Runtime-evaluated SQL correction experiments
- Benchmark comparison against deterministic SQL debugging tasks

### Downstream Use

- Fine-tuning initialization for enterprise SQL repair use cases
- Evaluation baseline for OpenEnv-style SQL agents

### Out-of-Scope / Not Recommended

- Autonomous execution against production databases without guardrails
- High-risk environments requiring strict SQL governance without additional review controls

## Training Details

### Training Data

Training signals are generated from deterministic OpenEnv SQL debugging tasks using reset/step interaction loops and execution-based grading.

### Training Procedure

| Step | Description |
|---|---|
| Session isolation | Every episode runs in isolated in-memory SQLite state |
| Task iteration | Query proposals are evaluated task-by-task under deterministic graders |
| GRPO objective | Relative ranking over generated candidates using execution-grounded reward |
| Artifact capture | Run metrics, reward traces, and charts are persisted and published |

### Key Training Hyperparameters (workflow-level)

| Hyperparameter area | Value / behavior |
|---|---|
| GRPO generations | Configured `>= 2` (runtime-safe default in launcher) |
| Reward composition | Correctness + efficiency + progress + schema bonus - penalties |
| Sampling controls | Temperature / top-p / completion length controlled in training scripts |

For script-level specifics, see:

- `ultimate_sota_training.py`
- `launch_job.py`

## Evaluation

### Metrics Snapshot

| Metric | Value |
|---|---:|
| Spider-style industry baseline | 48.2% |
| Qwen-7B base | 52.4% |
| RL agent headline | 78.5% |
| Performance leap view | 0.0% -> 25.0% |
| Eval artifact pass | 32-run |

### Benchmark Visuals

![Performance leap chart: baseline to RL-improved agent](https://md896-sql-debug-env.hf.space/static/chart-performance-leap.png)
![Comparison chart with reward shift](https://md896-sql-debug-env.hf.space/static/chart-comparison-shift.png)
![Spider-style benchmark headline chart](https://md896-sql-debug-env.hf.space/static/chart-spider-benchmark.png)

### Training / Proof Visuals

![Training reward curve over run steps](https://md896-sql-debug-env.hf.space/static/training_reward_curve_final.png)
![Dual-axis diagnostics across training](https://md896-sql-debug-env.hf.space/static/training_diagnostics_dual_axis_final.png)
![Baseline vs trained performance by task](https://md896-sql-debug-env.hf.space/static/baseline_vs_trained_by_task_final.png)
![Reward distribution shift after RL training](https://md896-sql-debug-env.hf.space/static/reward_distribution_shift_red_green_final.png)
![Cost versus performance curve](https://md896-sql-debug-env.hf.space/static/cost_vs_performance_final.png)

### Evidence Artifacts

- Sample rewards run folder: https://huggingface.co/spaces/md896/sql-debug-env/tree/main/artifacts/runs/20260426-064318-sample-rewards-32eval
- Earlier 32-eval pass folder: https://huggingface.co/spaces/md896/sql-debug-env/tree/main/artifacts/runs/20260426-060502-final-pass-32eval

## Bias, Risks, and Limitations

- SQL correctness can still degrade under unseen schemas/dialects.
- Benchmark-style gains do not guarantee equivalent production reliability.
- Model outputs should be reviewed before executing in sensitive environments.

## Recommendations

- Keep SQL execution sandboxed during evaluation.
- Use schema introspection + error inspection loops.
- Add reviewer/guardrail checks for risky query classes.
- Track run artifacts and compare against deterministic graders, not only manual inspection.

## How to Get Started

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "md896/sql-debug-agent-qwen25-05b-grpo-wandb-continue-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Fix this SQL query based on schema and error context: SELECT * FROM userss;"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Environmental Impact

This model was trained/evaluated across iterative cloud/local workflows. Exact carbon accounting is not yet logged in this card.

## Citation

If you use this work, cite the project repository and model page:

- Repo: https://github.com/mdayan8/sql-debug-env
- Model: https://huggingface.co/md896/sql-debug-agent-qwen25-05b-grpo-wandb-continue-v2

## Contact

- GitHub: https://github.com/mdayan8