Text Generation
MLX
Safetensors
English
Hungarian
bitnet
b1.58
ternary
quantization
apple-silicon
llm
sovereign
Instructions to use PeetPedro/quantal-ternary with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use PeetPedro/quantal-ternary with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("PeetPedro/quantal-ternary") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use PeetPedro/quantal-ternary with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "PeetPedro/quantal-ternary" --prompt "Once upon a time"
- Atomic Chat
Commit History
-
nightly: thresholded-ternary 168 matrices (best 2.1469)
Verified signature42f4da0
-
quantal ternary export from the SOTA run (epoch 21, val 0.5597)
Verified signatureTrained on a vast.ai H100 (mlx-cuda), early stop at epoch 21 of 40 after five epochs under the 0.02 min-delta. val_loss 2.8975 -> 0.5597. Source checkpoint quantal-long-final.safetensors, sha256 69af3001bcd273a55b7fef4a60199d4542a9049e749f882d8b299ae9e3140409, verified byte-identical between the training box and the object store before export. Note this is `final`, not the file the trainer names `best`: min-delta gates the save as well as the early stop, so `best` froze at epoch 16 (val 0.5772) while validation kept improving. final is 0.0175 better. Golden-logits gate against the Rust ternary runner: max abs delta 9.9e-05, argmax identical on both gate prompts, two orders of magnitude inside the 1e-2 tolerance.8dc4600
-
ULTRA part: matrices 010-049
Verified signaturedb0a20f -
refresh: winner ckpt masked val 3.2862 — 168 ternary matrices + embeddings + norms (deployed-forward QAT)
Verified signature7bbfe5d
-
Upload folder using huggingface_hub
Verified signature34e0c9b