Upload validation-selected step 1300 LoRA adapter

b91c06c verified about 1 month ago

2.4 kB

---
base_model: Qwen/Qwen3-4B-Instruct-2507
library_name: peft
license: apache-2.0
pipeline_tag: text-generation
tags:
  - text-to-sql
  - reinforcement-learning
  - rlvr
  - lora
  - bird
---
# ReViSQL-Qwen3-4B

LoRA adapter trained directly from
[`Qwen/Qwen3-4B-Instruct-2507`](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507)
with RLVR on expert-verified BIRD-Platinum data. This is the validation-selected
step 1,300 adapter from the
[`bird-text2sql-rl`](https://github.com/ximosss/bird-text2sql-rl) project.

## Results

| Dataset / decoding | Exact execution accuracy |
| --- | ---: |
| Arcwise-Plat-SQL, greedy | 376/498 = 75.50% |
| Arcwise-Plat-SQL, SC-16 | 414/498 = 83.13% |
| Arcwise-Plat, greedy | 404/498 = 81.12% |
| BIRD Full Dev, greedy | 961/1,534 = 62.65% |
| BIRD Mini-Dev, greedy | 298/500 = 59.60% |

Greedy uses temperature 0. SC-16 samples 16 candidates at temperature 1 and
selects the majority execution result. The model is evaluated as a SQL agent
with up to five rounds of read-only database interaction; see the source
repository for the exact prompt, environment, and evaluation contract.

## Loading the adapter

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "ximosss/ReViSQL-Qwen3-4B"
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
```

The adapter alone does not reproduce the benchmark protocol. Use the database
tool environment and prompts in the source repository for evaluation.

## Training details

- Objective: CISPO with execution, VeriEQL, and evidence-process rewards
- LoRA: rank 32, alpha 64; attention and MLP linear layers
- Learning rate: `1e-5`
- Rollout temperature: `0.8`
- Training batch / group size: 64 / 16

## References

- Yuxuan Zhu, Tengjun Jin, Yoojin Choi, and Daniel Kang. 2026.
  [Human-Level Text-to-SQL via Reinforcement Learning on Verified Data, Without Pipeline Engineering](https://arxiv.org/abs/2603.20004).
- Yuxuan Zhu, Tengjun Jin, Yoojin Choi, and Daniel Kang. 2026.
  [Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL](https://thinkingmachines.ai/news/putting-task-expertise-into-rl/).

	---
	base_model: Qwen/Qwen3-4B-Instruct-2507
	library_name: peft
	license: apache-2.0
	pipeline_tag: text-generation
	tags:
	- text-to-sql
	- reinforcement-learning
	- rlvr
	- lora
	- bird
	---

	# ReViSQL-Qwen3-4B

	LoRA adapter trained directly from
	[`Qwen/Qwen3-4B-Instruct-2507`](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507)
	with RLVR on expert-verified BIRD-Platinum data. This is the validation-selected
	step 1,300 adapter from the
	[`bird-text2sql-rl`](https://github.com/ximosss/bird-text2sql-rl) project.

	## Results

	\| Dataset / decoding \| Exact execution accuracy \|
	\| --- \| ---: \|
	\| Arcwise-Plat-SQL, greedy \| 376/498 = 75.50% \|
	\| Arcwise-Plat-SQL, SC-16 \| 414/498 = 83.13% \|
	\| Arcwise-Plat, greedy \| 404/498 = 81.12% \|
	\| BIRD Full Dev, greedy \| 961/1,534 = 62.65% \|
	\| BIRD Mini-Dev, greedy \| 298/500 = 59.60% \|

	Greedy uses temperature 0. SC-16 samples 16 candidates at temperature 1 and
	selects the majority execution result. The model is evaluated as a SQL agent
	with up to five rounds of read-only database interaction; see the source
	repository for the exact prompt, environment, and evaluation contract.

	## Loading the adapter

	```python
	from peft import PeftModel
	from transformers import AutoModelForCausalLM, AutoTokenizer

	base_model_id = "Qwen/Qwen3-4B-Instruct-2507"
	adapter_id = "ximosss/ReViSQL-Qwen3-4B"

	tokenizer = AutoTokenizer.from_pretrained(base_model_id)
	base_model = AutoModelForCausalLM.from_pretrained(
	base_model_id,
	torch_dtype="auto",
	device_map="auto",
	)
	model = PeftModel.from_pretrained(base_model, adapter_id)
	```

	The adapter alone does not reproduce the benchmark protocol. Use the database
	tool environment and prompts in the source repository for evaluation.

	## Training details

	- Objective: CISPO with execution, VeriEQL, and evidence-process rewards
	- LoRA: rank 32, alpha 64; attention and MLP linear layers
	- Learning rate: `1e-5`
	- Rollout temperature: `0.8`
	- Training batch / group size: 64 / 16

	## References

	- Yuxuan Zhu, Tengjun Jin, Yoojin Choi, and Daniel Kang. 2026.
	[Human-Level Text-to-SQL via Reinforcement Learning on Verified Data, Without Pipeline Engineering](https://arxiv.org/abs/2603.20004).
	- Yuxuan Zhu, Tengjun Jin, Yoojin Choi, and Daniel Kang. 2026.
	[Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL](https://thinkingmachines.ai/news/putting-task-expertise-into-rl/).