README.md

---
model_name: PhAI-IDE-9B
base_model: Qwen/Qwen3.5-9B
base_model_relation: finetune
datasets:
- AItonomy/ScienceIDE
library_name: transformers
pipeline_tag: image-text-to-text
license: apache-2.0
tags:
- phai-ide
- science
- code
- tool-use
- sft
- lora
- safetensors
---

# PhAI-IDE-9B

**PhAI-IDE-9B** is a supervised fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) for scientific coding and interaction with tools, trained using [ms-swift](https://github.com/modelscope/ms-swift). The release contains full BF16 weights with the final LoRA adapter merged.

The training data is sourced from [AItonomy/ScienceIDE](https://huggingface.co/datasets/AItonomy/ScienceIDE).

## Quick start

Use Transformers 5.16.1, PyTorch and Accelerate. The example loads the model, configuration and tokenizer directly from this repository.

```python
from transformers import AutoTokenizer, AutoModelForImageTextToText

model_id = "AItonomy/PhAI-IDE-9B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype="bfloat16", device_map="auto",
)
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain how to verify a numerical simulation."}],
    add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt",
).to(model.device)
output = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

BF16 weights occupy approximately 18.82 GB, plus runtime memory and KV cache.

## ScienceAccelBench performance

Across all 27 ScienceAccelBench environments, the task-pooled pass rate increases from **17.09% to 20.89% (+3.80 percentage points)** on **158 valid paired tasks**.

Task-held-out, localized scientific-code repair on familiar codebases, with original numerical verification. Pass rates (%); gains in percentage points.

| Environment | Tasks | Qwen3.5-9B | PhAI-IDE-9B | Gain (pp) |
| --- | ---: | ---: | ---: | ---: |
| laps | 16 | 31.25 | **50.00** | **+18.75** |
| mitgcm-biogeo | 8 | 0.00 | **12.50** | **+12.50** |
| pluto-rmhd-resrmhd | 7 | 0.00 | **28.57** | **+28.57** |

## Training procedure

ScienceIDE demonstrations were collected with **GPT-5.6-sol** and filtered using a **numerical-equivalence verifier**. They capture code inspection, tool use, and responses to execution feedback.

Training uses **ms-swift** supervised fine-tuning with **LoRA across trainable linear layers for three epochs**. The release merges the final checkpoint's adapter into the base model. Retained assistant targets provide the next-token training signal, while conversation history and tool observations provide context. The Codex-only data retains the native `exec` / `wait` interaction format. Heuristic target masking selects assistant actions for supervision while preserving the surrounding interaction history.

Fine-tuning uses text trajectories with the vision tower and aligner frozen. The merged export supports standard autoregressive Qwen3.5 generation.

| Setting | Value |
| --- | --- |
| Training dataset | Codex trajectories |
| Training examples / tasks | 4,567 segments / 564 tasks |
| Validation examples / tasks | 544 segments / 81 tasks |
| Train/validation task overlap | 0 |
| Training epochs | 3 |
| LoRA rank / alpha / dropout | 32 / 64 / 0.05 |
| Released weights | LoRA merged into BF16 Safetensors |
| Maximum training sequence length | 36,864 tokens |
| Learning rate / schedule | 2e-5 / cosine |
| Warmup ratio | 0.03 |
| Batch size per device / accumulation | 1 / 16 |
| LoRA targets | Linear modules in the unfrozen language model |
| Optimizer steps | 216 |
| Packing | Disabled |

Long trajectories are organized into segments. Source partition assignments are preserved, with no task identifiers shared between training and validation.

### Framework versions

The release was validated with the following environment.

| Component | Version |
| --- | --- |
| Python | 3.11 |
| ms-swift | 4.5.3 |
| Transformers | 5.16.1 |
| PyTorch | 2.6.0+cu124 |
| PEFT | 0.20.0 |
| Datasets | 4.8.4 |
| Tokenizers | 0.23.2 |
| Accelerate | 1.14.0 |

## Citations

The training framework is [ms-swift](https://github.com/modelscope/ms-swift), and adapter training uses [LoRA](https://arxiv.org/abs/2106.09685).

```bibtex
@misc{zhao2024swift,
  title = {SWIFT: A Scalable lightWeight Infrastructure for Fine-Tuning},
  author = {Yuze Zhao and others},
  year = {2024},
  eprint = {2408.05517},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.05517}
}

@misc{hu2021lora,
  title = {LoRA: Low-Rank Adaptation of Large Language Models},
  author = {Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen},
  year = {2021},
  eprint = {2106.09685},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.09685}
}
```