An execution-grounded model for evolutionary machine learning engineering and AutoResearch
š Paper ā¢ š Project ā¢ š» Code ā¢ š¤ Models ā¢ š§© Tasks ā¢ š SFT Traces
## Introduction **Frontis-MA1-30B** is the companion 30B open-weight research model released with **OpenMLE**, an execution-grounded system for studying meta-evolution in machine learning engineering (MLE). Starting from [Qwen3-30B-A3B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507), we apply execution-grounded supervised fine-tuning and reinforcement learning so that the model learns four reusable program-transformation operators: - **Draft** ā create a complete initial solution; - **Improve** ā refine a parent program using its score and execution evidence; - **Debug** ā repair invalid or failing code; - **Crossover** ā recombine useful elements from two parent solutions. At inference time, OpenMLE-Evo composes these operators into a long-horizon search loop. Frontis-MA1 is therefore both a product of the OpenMLE training stack and the model that drives its evolutionary search. ## Highlights - **Execution-grounded post-training.** SFT examples and RL rewards are constructed from programs executed in isolated MLE sandboxes. - **Trainable evolutionary operators.** Draft, Improve, Debug, and Crossover provide a shared action space across training and test-time search. - **Linked research release.** The model card connects the weights to the OpenMLE training code, search protocol, task environments, and evaluation boundary. - **Strong controlled model gain.** Under the same OpenMLE-Evo harness on MLE-Bench Lite, Frontis-MA1-30B improves Medal Average from **34.85% to 53.03%** and Human Rank from **0.5573 to 0.7055** over its base model. > The reported scores measure a **modelāharness system**, not standalone one-shot generation. Reproduction requires the same OpenMLE-Evo search configuration, task environments, sandbox budget, and evaluation protocol. ## Model family | Model | Base model | Release format | Status | | --- | --- | --- | --- | | **Frontis-MA1-30B** | Qwen3-30B-A3B-Thinking-2507 | BF16 Transformers | This repository | | [Frontis-MA1-30B-GGUF](https://huggingface.co/FrontisAI/Frontis-MA1-30B-GGUF) | Frontis-MA1-30B | Q4_K_M GGUF | Local deployment | | [Frontis-MA1-35B](https://huggingface.co/FrontisAI/Frontis-MA1-35B) | Qwen3.6-35B-A3B | BF16 Transformers | Companion release | ## Model details | Field | Value | | --- | --- | | Model family | Frontis-MA1 | | Architecture | Mixture-of-Experts causal language model | | Parameters | 30.5B total, 3.3B activated | | Layers | 48 | | Experts | 128 total, 8 activated per token | | Weight precision | BF16 | | Native configuration context | 262,144 tokens | | Post-training SFT cutoff | 32,768 tokens | | License | CC BY-NC 4.0 | The 262,144-token context length is inherited from the base-model configuration. OpenMLE post-training used a 32,768-token SFT cutoff; substantially longer contexts have not yet been validated to the same standard as the reported experiments. ## Training Frontis-MA1-30B uses full-parameter supervised fine-tuning followed by reinforcement learning from executable task feedback. ### Supervised fine-tuning - 26,259 released execution-grounded examples: 17,245 full responses and 9,014 trajectory steps; - full-solution and trajectory-step supervision; - BF16 full-parameter training on 8 NVIDIA H200 GPUs; - global batch size 128; - learning rate `3e-5` with cosine decay and 0.1 warmup fraction; - three epochs; - Qwen3-compatible thinking-loss masking. ### Reinforcement learning - online generation and evaluation in isolated task sandboxes; - operator mixture: Draft 0.50, Improve 0.17, Debug 0.17, Crossover 0.16; - 16 prompts per rollout and 16 samples per prompt; - maximum response length 24,576 tokens; - GSPO with execution-grounded reward post-processing; - Adam optimizer with learning rate `1e-6`. See the paper and [OpenRSI code](https://github.com/FrontisAI/OpenRSI) for the complete data-construction, training, search, and evaluation protocol.
Paper framework figure. Frontis-MA1 learns the same four atomic operators used by OpenMLE-Evo, first from executable SFT rollouts and then from online RL execution feedback.
## Benchmark results The headline numbers are **modelāharness system results**, not standalone one-shot scores. The controlled comparison holds the OpenMLE-Evo harness fixed and changes only the model from the Qwen base to Frontis-MA1-30B. ### MLE-Bench Lite protocol | Item | Setting | | --- | --- | | Benchmark | Official 22-task MLE-Bench Lite split | | Independent runs | 3 per OpenMLE-Evo configuration, unless stated otherwise | | Sandbox budget | 12 hours on 0.5 NVIDIA RTX 4090 equivalent per task | | Equivalent compute | 6 RTX 4090 GPU-hours per task | | Execution | Generated programs are run and scored in isolated task sandboxes | | OpenMLE-Evo-Max | MLE-Bench-disjoint priors plus asynchronous multi-GPU search | | Budget comparison | OpenMLE-Evo-Max keeps the same total sandbox-compute budget | The paper reports: - **Valid Rate:** mean number of the 22 tasks with a valid submission; - **Medal Average:** mean fraction of tasks receiving any Kaggle medal; - **Human Rank:** mean fraction of human leaderboard participants surpassed by the selected submission. All three metrics are higher-is-better. ### Main results from the paper
Current paper main-results figure. The left panel shows all completed modelāharness results on MLE-Bench Lite; the right panel compares model size with Medal Average. Orange denotes Frontis-MA1 with OpenMLE-Evo, cyan denotes other models with OpenMLE-Evo, hatching denotes OpenMLE-Evo-Max, and gray denotes general-purpose coding harnesses. The paper highlights the primary 35B model in this figure and reports the companion 30B results in its main table.
### Controlled 30B comparison | Model | Harness | Valid Rate ā | Medal Average ā | Human Rank ā | | --- | --- | ---: | ---: | ---: | | Qwen3-30B-A3B-Thinking-2507 | OpenMLE-Evo | 17.33/22 | 34.85% | 0.5573 | | **Frontis-MA1-30B** | OpenMLE-Evo | **21.67/22** | **53.03%** | **0.7055** | | **Frontis-MA1-30B** | OpenMLE-Evo-Max | **22.00/22** | **66.67%** | **0.8053** | | Comparison | Ī Valid Rate | Ī Medal Average | Ī Human Rank | What it measures | | --- | ---: | ---: | ---: | --- | | Frontis-MA1-30B vs. Qwen base, OpenMLE-Evo fixed | +4.34 tasks | +18.18 pp | +0.1482 | Post-training gain under a fixed harness | | OpenMLE-Evo-Max vs. OpenMLE-Evo, Frontis-MA1-30B fixed | +0.33 task | +13.64 pp | +0.0998 | Additional system-level search gain | The fixed-harness comparison is the primary evidence for model improvement. OpenMLE-Evo-Max also changes the search system, so its result must be read as an end-to-end modelāharness score rather than a pure model gain. These experiments do not establish a standalone ranking outside the OpenMLE evaluation protocol. The paper does not provide a 30B-only result figure, so this model card keeps the original paper values as a searchable table rather than introducing a derived visualization. ### Public-weight model comparison under OpenMLE-Evo The following rows use the same OpenMLE-Evo harness, benchmark split, and sandbox-compute budget. This is the closest available comparison with strong public-weight models, although model scale, pretraining data, and architecture still differ. | Public-weight model | Valid Rate ā | Medal Average ā | Human Rank ā | | --- | ---: | ---: | ---: | | [Kimi K2.6](https://huggingface.co/moonshotai/Kimi-K2.6) | 21.67/22 | **66.67%** | 0.7859 | | [GLM-5.2](https://huggingface.co/zai-org/GLM-5.2) | 19.67/22 | 62.12% | 0.7069 | | [Frontis-MA1-35B](https://huggingface.co/FrontisAI/Frontis-MA1-35B) | 21.67/22 | 60.61% | 0.7647 | | [MiniMax M3](https://huggingface.co/MiniMaxAI/MiniMax-M3) | **22.00/22** | 59.09% | **0.7994** | | **Frontis-MA1-30B** | 21.67/22 | 53.03% | 0.7055 | | [DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) | 21.33/22 | 51.52% | 0.6957 | | [Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) | 19.67/22 | 39.39% | 0.5828 | | [Qwen3-30B-A3B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507) | 17.33/22 | 34.85% | 0.5573 | These are modelāharness results rather than standalone one-shot model scores. The full paper figure above additionally includes OpenMLE-Evo-Max and general-purpose coding-agent systems, whose harnesses are not directly interchangeable with this fixed-harness table. ## Usage ### Transformers ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "FrontisAI/Frontis-MA1-30B" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto", ) messages = [ { "role": "user", "content": ( "Build a strong tabular classification baseline and explain " "the validation design." ), } ] inputs = tokenizer.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=2048) completion = outputs[0, inputs["input_ids"].shape[1]:] print(tokenizer.decode(completion, skip_special_tokens=True)) ``` ### vLLM ```bash vllm serve FrontisAI/Frontis-MA1-30B \ --served-model-name Frontis-MA1-30B \ --tensor-parallel-size 8 \ --max-model-len 32768 \ --enable-reasoning \ --reasoning-parser deepseek_r1 ``` ## Intended use Frontis-MA1-30B is intended for: - research on machine learning engineering and AutoResearch agents; - generation and iterative improvement of ML experiment code; - execution-grounded program search; - use with OpenMLE-Evo or compatible sandboxed harnesses. Generated code may be incorrect, insecure, destructive, or expensive to execute. Run it inside an isolated environment with explicit CPU, memory, GPU, network, filesystem, and time limits. ## Limitations - The main evaluation centers on MLE-Bench Lite and does not establish general performance across all software-engineering or scientific-research tasks. - The paper results depend on an external evolutionary search harness and sandbox feedback. - Execution reward does not fully measure maintainability, robustness, security, or responsible data handling. - Context lengths beyond the post-training cutoff require additional validation. - Outputs may contain reasoning traces and should be handled according to downstream requirements. ## Release artifacts This repository is the canonical BF16 Transformers release. The [Frontis-MA1-30B-GGUF](https://huggingface.co/FrontisAI/Frontis-MA1-30B-GGUF) repository provides the Q4_K_M local-deployment derivative. ## Paper and citation **Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering** Preprint: [arXiv:2607.28568](https://arxiv.org/abs/2607.28568). ## License Original Frontis-MA1 material is released under [CC BY-NC 4.0](LICENSE) for attribution-required, non-commercial use. Commercial use is not granted. The upstream Qwen Apache License 2.0 notice is preserved in [LICENSE-UPSTREAM-APACHE-2.0](LICENSE-UPSTREAM-APACHE-2.0) and [NOTICE](NOTICE). The license in this repository applies to the model weights, configuration files, and model documentation distributed here. Training and evaluation datasets and third-party software remain subject to their respective terms. ## Acknowledgements We thank the Qwen team and the open-source communities behind Transformers, SLIME, Ray, Megatron-LM, SGLang, MLE-Bench, and the broader executable MLE research ecosystem.