Instructions to use ko120/MLE-World-Qwen3.5-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ko120/MLE-World-Qwen3.5-4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ko120/MLE-World-Qwen3.5-4B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("ko120/MLE-World-Qwen3.5-4B")
model = AutoModelForMultimodalLM.from_pretrained("ko120/MLE-World-Qwen3.5-4B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ko120/MLE-World-Qwen3.5-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ko120/MLE-World-Qwen3.5-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ko120/MLE-World-Qwen3.5-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ko120/MLE-World-Qwen3.5-4B

SGLang

How to use ko120/MLE-World-Qwen3.5-4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ko120/MLE-World-Qwen3.5-4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ko120/MLE-World-Qwen3.5-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ko120/MLE-World-Qwen3.5-4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ko120/MLE-World-Qwen3.5-4B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ko120/MLE-World-Qwen3.5-4B with Docker Model Runner:
```
docker model run hf.co/ko120/MLE-World-Qwen3.5-4B
```

MLE World — Qwen3.5-4B

The execution world model used in the Qwen3.5-9B stage-policy experiments recorded in stage_policy_qwen35_9b_seven_arm_8h_plan_20261011.md. This is checkpoint 6685, epoch 5.0, from the mle_qwen3_4b_sft_world_9_3_exc_token_h100_adamw_h200 training run. The saved serving audit for the adopted main experiment confirms this exact checkpoint.

The model is a full supervised fine-tune of Qwen/Qwen3.5-4B on MLE World: 497,056 training examples and 1,700 held-out examples across 17 MLE-bench tasks.

Behavior

Given a Python ML program, installed-package environment, data overview, and runtime budget, the model predicts the exception type, terminal output, validation metric, and whether a submission file exists. It uses the env_v1_reorder_excfirst_token prompt format:

<exc_type>EXC_None</exc_type>
<term_out>...</term_out>
<val_metric>...</val_metric>
<submission_exists>true</submission_exists>

Use the system and human messages from the dataset for the exact input format. The matching tokenizer contains 82 ordinary exception tokens; these remain visible when decoding with skip_special_tokens=True.

Predictions estimate program behavior; real execution and grading were used to validate selected programs in the experiments.

Load for text inference

The checkpoint was saved with Transformers 5.9.0. Use a Transformers version supporting Qwen3_5ForConditionalGeneration.

import torch
from datasets import load_dataset
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

repo = "ko120/MLE-World-Qwen3.5-4B"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    repo, dtype=torch.bfloat16, device_map="auto"
)
example = next(iter(load_dataset(
    "ko120/MLE-World", split="validation", streaming=True
)))
messages = [
    {"role": "system", "content": example["system"]},
    {"role": "user", "content": example["conversations"][0]["value"]},
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, do_sample=False, max_new_tokens=8192, use_cache=True)
print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Experiment inference used temperature 0, top_p=1, top_k=-1, min_p=0, repetition penalty 1, no presence/frequency penalties, seed 0, a 32,768-token serving context, and an 8,192-token output cap.

Training

Setting	Value
Method	Full SFT; vision tower and multimodal projector frozen
Training epochs / final step	5 / 6685
Learning rate	1e-5
Optimizer	Fused AdamW
Scheduler / warmup	Cosine / 5%
Packed sequence length	12,288 tokens
Packing	Enabled, with neat packing
Effective batch	128 packed sequences across 8 GPUs
Final evaluation loss	0.0814168155

The evaluation loss is the training framework's token loss on the held-out corpus, with its packing and length processing. It is not execution-prediction accuracy.

training_config.original.yaml preserves the original training configuration and its local paths. Replace the model, dataset, and output paths for a new environment. Training logs and results are retained under training/.

Artifact fidelity

The safetensors weights and tokenizer are copied from checkpoint 6685 without numerical conversion. The source stores 426 FP32 tensors and 297 BF16 tensors; its configuration declares BF16 for model loading. provenance.json records SHA-256 hashes, source checkpoint identity, tensor counts, and the experiment reference.

The model is distributed under the base model's Apache 2.0 license, included in LICENSE.