MachineLearningLM
This repository contains the model presented in the paper MachineLearningLM: Scaling Many-shot In-context Learning via Continued Pretraining.
Model Summary
Can LLMs learn from 1,000 in-context examples?
Introducing MachineLearningLM 🧪📊 — a model continuously pretrained on millions of synthetic tabular ML tasks, enabling robust many-shot in-context learning.
📈 Scales from 8 to 1,024 examples
📈 ~15% improvement on unseen tabular tasks compared to o3-mini / GPT-5-mini / Qwen-2.5-7B-Instruct
🌲 Random-Forest–level numerical modeling robustness
🧠 MMLU score: 75.4%
📄 Read the paper: https://huggingface.co/papers/2509.06806
GitHub: https://github.com/HaoAreYuDong/MachineLearningLM
Evaluation and Validation
We have developed an automated evaluation framework — simply configure the parameters to easily perform validation and evaluation. The code is now open-sourced at our GitHub repository.
Quick Start
pip install -r requirements.txt
python ./src/evaluation/model_pred/dl_model_pred.py \
--input_dir ./demo_input.jsonl \
--output_dir ./demo_output.jsonl \
--model_name MachineLearningLM/MachineLearningLM-7B-v1
Pipeline
# modify the evaluate_parameters.sh file
source evaluate_parameters.sh
# Option 1 End-to-End Pipeline
./scripts/evaluate_pipeline.sh
# Option 2 Parallel Processing
./scripts/multi_process/data_prep.sh
./scripts/multi_process/prompt_gen.sh # For deep learning only
./scripts/multi_process/model_pred.sh
./scripts/multi_process/evaluation.sh
./scripts/multi_process/report.sh
# Option3 Sequential Processing
./scripts/single_process/data_prep.sh
./scripts/single_process/prompt_gen.sh # For deep learning only
./scripts/single_process/model_pred.sh
./scripts/single_process/evaluation.sh
./scripts/single_process/report.sh
For more usage details, please visit our GitHub.
Quants of Checkpoints
https://huggingface.co/QuantFactory/MachineLearningLM-7B-v1-GGUF
Tabicl Evaluation
This part of the code needs to run in an environment with the tabicl and openpyxl libraries installed.
The evaluation code for tabicl is placed separately in the ./src/evaluation/tabicl_evaluate.py file. Use ./scripts/tabicl_evaluate.sh to obtain the evaluation results for tabicl.
Use --datasets to specify the datasets to be evaluated, and --sample_sizes to indicate the number of shots.
If multiple datasets need to be evaluated, separate them with spaces. To evaluate all CSV files in the input folder, use all.
Prior_data
MachineLearningLM uses the code from tabicl to generate prior data.
Use ./scripts/generate_data.sh to generate the prior data. It generates the corresponding .pt and .csv files, and normalizes the feature values in the CSV files to the range of 0–999, as we did in the paper.
Parameter Introduction（refer to the comments in the file tabicl\src\tabicl\prior\dataset.py）
Data Scale & Structure
| Parameter | Type | Description |
|---|---|---|
min_features | int | Minimum number of features per dataset |
max_features | int | Maximum number of features per dataset |
max_classes | int | Maximum number of target classes |
min_seq_len | int | Minimum samples per dataset. Uses max_seq_len if None |
max_seq_len | int | Maximum samples per dataset （Not Include） |
Batch Configuration
| Parameter | Type | Description |
|---|---|---|
batch_size | int | Total number of datasets to generate per batch |
batch_size_per_gp | int | Number of datasets per group (shared characteristics) |
batch_size_per_subgp | int | Number of datasets per subgroup (similar causal structures). Defaults to batch_size_per_gp if None |
Sequence Length Control
| Parameter | Type | Description |
|---|---|---|
log_seq_len | bool | Sample sequence length from log-uniform distribution if True |
seq_len_per_gp | bool | Sample sequence length per group (enables variable-sized datasets) |
replay_small | bool | Occasionally sample smaller sequences for model robustness |
Train-Test Split
| Parameter | Type | Description |
|---|---|---|
min_train_size | int/float | Start position/ratio for train split (int: absolute, float: fractional) |
max_train_size | int/float | End position/ratio for train split (int: absolute, float: fractional) |
Generation Method
| Parameter | Type | Description |
|---|---|---|
prior_type | str | Prior type: 'mlp_scm', 'tree_scm', or 'mix_scm' (random selection) |
fixed_hp | dict | Fixed structural configuration parameters |
sampled_hp | dict | Parameters sampled during generation |
Computation Settings
| Parameter | Type | Description |
|---|---|---|
n_jobs | int | Number of parallel jobs (-1 = use all processors) |
num_threads_per_generate | int | Number of threads per generation job |
device | str | Computation device ('cpu' or 'cuda') |
Train
MachineLearningLM uses the LLaMA-Factory framework for training.
Training Environment Configuration
cd ./third_party/LLaMA-Factory
pip install -e ".[torch,metrics]" --no-build-isolation
pip install wandb
Use ./scripts/train.sh for training.
Project Structure
MachineLearningLM/
├──src/
| ├──evaluation/
│ │ ├── data_prep/ # Data preprocessing and chunking utilities
│ │ ├── prompt_gen/ # Prompt generation for deep learning models
│ │ ├── model_pred/ # Model inference (ML and DL prediction engines)
│ │ ├── result_proc/ # 5-layer evaluation architecture and metrics processing
│ │ ├── zero_summary/ # Result summarization and report generation
│ │ └── tabicl_evaluate.py
│ └──prior_data
│ └── pt_to_csv.py
├── scripts/
│ ├── single_process/ # Sequential execution shell scripts
│ ├── multi_process/ # Parallel execution shell scripts (with _mp suffix)
│ ├── evaluate_parameters.sh # Global parameter configuration
| ├── evaluate_pipeline.sh # automated pipeline
| ├── generate_data.sh
| ├── tabicl_evaluate.sh
| └── train.sh
├── datahub_inputs/
│ ├── data_demo/ # Demo datasets for testing
│ └── data_raw/ # Raw input datasets
├── third_party/
│ ├── tabicl/
│ └── LLaMA-Factory/
├── requirements.txt # Python dependencies for Evaluation Framework
├── README.md
├── README_zh.md
├── THIRD_PARTY_NOTICES.md
└── LICENSE