Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Training Large Language Model to Reason in a Continuous Latent Space
| Date | Stars |
|---|---|
| 2026-07-31 | 1676 |
| 2026-08-06 | 1676 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Coconut The code base is the official implementation of [Training Large Language Models to Reason in a Continuous Latent Space](https://arxiv.org/abs/2412.06769).  ## Getting Started Clone repo: ``` git clone [email protected]:facebookresearch/coconut.git cd coconut ``` Setup environment: ``` conda create --name coconut python=3.12 conda activate coconut pip install -r requirements.txt ``` The code relies on [wandb](https://wandb.ai/site/) for logging. Please log in your wandb account following this [document](https://docs.wandb.ai/ref/cli/wandb-login/) before running any experiments. ## Data The data for training and evaluation should be presented as a json file like below: ```python [ { "question": "...", "answer": "...", "steps": ["...", "...", ...] }, ... ] ``` The file should contain a list of data points. Each data point is composed of a question (str), an answer (str), and a list of steps (str), where each of them is a string. For example, you can download and process the [GSM8K](https://arxiv.org/abs/2110.14168) dataset (with [augmented training and validation sets](https://github.com/da03/Internalize_CoT_Step_by_Step/tree/e06a32ee5e4cd117171daeb4755d2a97ece62761/data/gsm8k)) by running: ```bash bash preprocessing/gsm_icot.bash ``` ## Arguments The configuration of a run should be specified in a yaml file (an example can be found [here](args/gsm_coconut.yaml)). - **General settings** - **project**: Project name for wandb - **save_path**: Your path to store the checkpoints - **only_eval**: If true, only load a model and test on the data from `val_path` (must used along with `load_model_path`). Otherwise, train the model on `train_path` and test on `val_path` after every epoch. - **Method** - **coconut**: Train coconut model - **cot**: Train cot model - **no_thoughts**: Train coconut (w/o thought) model - **no_cot**: Train no-cot model - **Training settings** - **c_thought**: Number of continuous thoughts for each reasoning step - **epochs_per_stage**: Number of epochs for every training stage - **max_latent_stage**: The maximum number of training stages (in addition to the initial stage) - **pad_latent_to_max**: If the number of reasoning steps is fewer than the index of current training stage, pad the number of continuous thoughts. - **save_only_improve**: Save the model only when there the best validation accuracy is updated. Recommended to set `False` for Coconut model training, because otherwise the checkpoints in the last stage might now get saved. - **uniform_prob**: The probability to mix data from other stages. 0 for standard experiment, 0.3 for analysis experiment. - **model_id**: Huggingface model id to load as the initialization, e.g., `openai-community/gpt2` - **load_model_path**: The path to a checkpoint to load. Used in two cases: (1) for evaluation (2) to initialize coconut from a CoT-tuned model. - **seed**: Random seed. - **resume**: The epoch to resume. Can be used when we want to skip the initial training stages. - **bf16**: Whether to use bf16 training. - **train_path**: Path to the training set. - **val_path**: Path to the validation or test set (depending on `only_eval`) - **reset_optimizer**: Whether to reset the optimizer when swtiching training stages. - **batch_size_training**: Batch size to train the model per GPU. - **debug**: If true, there is no wandb and model saving. A subset of data will be used. - **gradient_accumulation_steps**: Gradient accumulation steps - **num_epochs**: Maximum training epoches. - **lr**: Learning rate - **weight_decay**: Weight decay ## Training Run the following commands (replacing `N_GPUS` and `PATH_TO_ARGS`): ``` torchrun --nnodes 1 --nproc_per_node N_GPUS run.py PATH_TO_ARGS ``` ## Reproducing Experiments Here we provide instructions to reproduce our experiments in the paper. All the commands below assume 4 * A100 (80GB) GPUs. You may c
Excerpt of 6,888 characters
Read on GitHub
Shibo Hao · UC San Diego · United States
3
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:5d6e83b3db3589c0, llm:Repository description: 'Training Large Language Model to Reason in a Continuous Latent Space' (facebookresearch/coconut). Python project for LLM training and reasoning in latent space.
matched fp:5d6e83b3db3589c0, llm:Repository description: 'Training Large Language Model to Reason in a Continuous Latent Space' (facebookresearch/coconut). Python project for LLM training and reasoning in latent space.
matched fp:5d6e83b3db3589c0, llm:Repository description: 'Training Large Language Model to Reason in a Continuous Latent Space' (facebookresearch/coconut). Python project for LLM training and reasoning in latent space.