Rewrite model card in plain language

README.md

@@ -16,9 +16,11 @@ tags:
 
 # Sol Lassi 600K
 
-Sol Lassi is a 600,000-parameter language model trained from scratch on 1 billion token exposures. It uses M-SimOW and runs with MLX on Apple Silicon.
+Sol Lassi is a 600,000-parameter experiment in training a language model with M-SimOW. It runs with MLX on Apple Silicon. Training started from scratch and used 1 billion token exposures.
 
-## Model
+Lassi is a base model for text completion. Its size makes it useful for optimizer experiments, but completions can repeat, lose coherence, or give incorrect answers. It hasn't been instruction-tuned.
+
+## Model details
 
 | Setting | Value |
 |---|---|
@@ -33,9 +35,9 @@ Sol Lassi is a 600,000-parameter language model trained from scratch on 1 billio
 | Training exposures | 1,000,000,000 |
 | Weights | MLX NPZ |
 
-## Use
+## Run a completion
 
-Download the repository, install its dependencies, and add its directory to the Python path before importing the loader:
+Download the repository and install the dependencies in `requirements.txt`. Add the downloaded directory to Python's path before importing the loader:
 
 ```bash
 pip install -r requirements.txt
@@ -53,20 +55,18 @@ model, tokenizer = load_model(model_dir)
 print(generate(model, tokenizer, "A tiny language model can", max_new_tokens=48))
 ```
 
-## Training
-
-The model used a balanced FinePhrase stream drawn from the `faq`, `math`, `table`, and `tutorial` configurations. The FinePhrase-balanced-500m-2k-v2 manifest contains 500M stream tokens, which produced 1B training exposures.
+## Data and optimizer settings
 
-Training used 128-token sequences, a batch size of 32, and seed 7. M-SimOW used momentum beta 0.8 and decoupled weight decay 0.1. The learning rate was 0.006 for the first 500M exposures, then 0.003 for the remaining 500M.
+Training used a balanced FinePhrase stream from the `faq`, `math`, `table`, and `tutorial` configurations. The FinePhrase-balanced-500m-2k-v2 manifest has 500M stream tokens; the run accumulated 1B exposures.
 
-## Evaluation and limits
+The sequence length was 128 tokens, with a batch size of 32 and seed 7. M-SimOW used momentum beta 0.8 and decoupled weight decay 0.1. The first 500M exposures used a learning rate of 0.006, followed by 0.003 for the remaining 500M.
 
-No Open SLM, ArithMark, or Intelligence Index evaluation is included. The training-loss logs measure individual minibatches and are not held-out scores.
+## What has been measured
 
-Lassi is a base-model and optimizer experiment. At this size, completions can be incoherent, repetitive, or incorrect. It has not been instruction-tuned.
+This repository doesn't include an Open SLM, ArithMark, or Intelligence Index evaluation. The loss logs record individual training minibatches, so they don't give a held-out score.
 
 ## Files and license
 
-`model.npz` contains the weights. `modeling_sol_lassi.py` provides the loader and generator, which depend on `keystone_mlx/`. The repository also includes the tokenizer, architecture configuration, and `training_state.json`.
+The weights are in `model.npz`. `modeling_sol_lassi.py` supplies the loader and generator and depends on `keystone_mlx/`. The repository also has the tokenizer and architecture configuration, with run metadata in `training_state.json`.
 
-The model is licensed under [CC BY 4.0](LICENSE). FinePhrase and its sources retain their own terms.
+The model uses [CC BY 4.0](LICENSE). FinePhrase and its source datasets keep their own terms.