Wisp-5M

Wisp-5M is a tiny decoder-only language model from the Wisp family, trained from scratch on 3.00B tokens of English web text, code and math. It uses a plain LLaMA architecture, so it loads with transformers (LlamaForCausalLM) with no custom code.

It is a base model (not instruction-tuned). Its size makes it useful for research, education, speculative decoding drafts, on-device experiments and as a starting point for fine-tuning. Don't expect factual accuracy.

The Wisp family

All models share the same tokenizer and saw the same data in the same order: one fixed global shuffle of 2048-token windows. That makes the family directly comparable across sizes.

Architecture

• LLaMA-style decoder: pre-norm RMSNorm, rotary position embeddings (θ = 10,000), SwiGLU MLP, grouped-query attention, no biases, tied input/output embeddings
• Deep-and-thin shapes (following MobileLLM), which work better than wide-and-shallow at this scale
• Context length: 2048 tokens
• Tokenizer: Wisp byte-level BPE, 8,192 tokens, trained on the same data mix; digits are split individually (better for arithmetic); <|endoftext|> is used as BOS/EOS and document separator. Compression is ~3.8 characters/token on English web text and ~3.0 on code.

Training

• Compute cost: ~US$2.28 on Hugging Face Jobs

Data mix

An 8B-token pool was pretokenized with the mix below and shuffled once as 2048-token windows. Each model trained on the first 3.00B tokens of that shuffle, which keeps the same proportions.

Evaluation

Held-out cross-entropy (nats/token, lower is better) on documents never seen in training:

The average is unweighted across the four sources. Code is much more predictable than prose, so it sits below the training loss, which follows the 45/35/12/8 training mix. Weighted by that mix, the held-out loss is 3.080.

Zero-shot benchmarks

lm-evaluation-harness, zero-shot; acc_norm for HellaSwag/ARC/PIQA/OpenBookQA/SciQ, acc for WinoGrande/LAMBADA. Pythia-70M (70M params, 300B tokens of the Pile) is shown as a reference from EleutherAI's published evaluations.

Greedy samples right after training:

• 'The meaning of life is the same as the meaning of life. The meaning of life is the same as the meaning of life.\nThe meaning of life is the same as the meaning of life. It means that we know what we are doing and how we can do'
• 'Once upon a time, the world was not as much of an issue. The world had to be more than just a single day and a week.\nThe world was very different from what we have today. We are all in our own homes. It is a great'
• 'def is_prime(n): 10\n# 2. The number of digits in the first digit is 3, and the last digit is 4.\n# 3. The number of digits in the second digit is 5, and the last digit'
• 'The derivative of x^2 is the product of its value. The derivative of y^2 is the product of its value.\n\n## 1. What are the derivatives?\n\n### 3. How do you find the derivative of x^2?\n\n### '

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "DedeProGames/Wisp-5M"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

inputs = tok("The water cycle is", return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=60, do_sample=True, temperature=0.8, top_p=0.95,
                     repetition_penalty=1.1)
print(tok.decode(out[0], skip_special_tokens=True))

The tokenizer prepends <|endoftext|> as BOS, which matches how documents were laid out during training.

Limitations

A model this small has very limited knowledge and reasoning. It will produce fluent-looking but often wrong or incoherent text, and it can reproduce biases present in web data. It is English-centric and was not aligned or safety-tuned. Don't use it for anything where correctness matters.

License

Apache-2.0 for the model weights. Training data licenses: FineWeb-Edu, FineMath and The Stack v3 are ODC-By; DCLM-Baseline is CC-BY-4.0. Code in The Stack v3 comes with its original licenses.