Mnemonic
Small chat models where every part (data pipeline, tokenizer, GPU training, quantization and inference) is hand-written assembly: x86-64 on the CPU, PTX on the GPU, WebAssembly in the browser.
Code: https://github.com/Supergoatscriptguy/Mnemonic
| file | model | weights | size |
|---|---|---|---|
mnemonic-755m-q8.mnm |
755M | int8, one scale per row | 722 MB |
mnemonic-755m-q4.mnm |
755M | int4, one scale per group of 32 | 450 MB |
mnemonic-300m-q8.mnm |
300M | int8, one scale per row | 291 MB |
mnemonic-300m-q4.mnm |
300M | int4, one scale per group of 32 | 181 MB |
mnemonic-q8.mnm |
126M | int8, one scale per row | 121 MB |
mnemonic-q4.mnm |
126M | int4, one scale per group of 32 | 75 MB |
mnemonic-math-q8.mnm |
Math (300M) | int8, one scale per row | 291 MB |
mnemonic-math-q4.mnm |
Math (300M) | int4, one scale per group of 32 | 181 MB |
Models. All three are Llama-style decoders (RoPE, RMSNorm, SwiGLU, grouped-query attention, tied embeddings, no biases), with a byte-level BPE vocab of 32768.
- 755M: 28 layers, d_model 1536, 24 query / 8 key-value heads, ffn 4096, context 2048.
- 300M: 24 layers, d_model 1024, 16 query / 4 key-value heads, ffn 2816, context 2048.
- 126M: 16 layers, d_model 768, 12 query / 4 key-value heads, ffn 2048, context 1024.
Training. All three were trained on one RTX 5070 Ti at home.
- 755M: pretrained on 15B tokens of FineWeb-Edu with the matmuls in MXFP8 (8-bit floats with a shared scale per block of 32) and bf16 AdamW moments, which is what made it fit in 16 GB. About 8.3 days, validation loss 2.63 (the 300M gets 2.84 on the same text). Then fine-tuned for chat in one pass of smol-smoltalk plus an identity set, at 2048 context (5.1 hours).
- 300M: pretrained on 10B tokens of FineWeb-Edu (about 90 hours, validation loss 2.84). Then fine-tuned for chat the same way (3.4 hours).
- 126M: pretrained on 5B tokens (19 hours, validation loss 3.10), then fine-tuned for chat on smol-smoltalk plus a smaller identity set.
Mnemonic Math. A separate model that only does math: the 300M, fine-tuned on about 22 million generated problems (4.6B tokens, about 40 hours) that are worked out one small step at a time, in 38 kinds: column arithmetic and long division, fractions, decimals, percents, ratios, powers, square roots, prime factorization, gcd and lcm, order of operations, linear and quadratic equations, systems, inequalities, absolute value, expanding and factoring, polynomial division, lines, sequences, statistics, geometry, unit conversions, money (discounts, tax, tips, change, interest), percent change, probability, equations like x² = 49, logarithms, derivatives of polynomials and word problems. It gets 99.6% of 3,085 held-out problems right (all but one of the misses are wording, not math), 697 of 700 of the 7 newest kinds, and 196 of 201 questions typed the way people type them. The int4 file scores the same on the held-out problems and gets 198 of the typed ones. Run it greedy: bin\chat model=models\mnemonic-math-q8.mnm temp=0 max=2000, or ollama run supergoatscriptguy/mnemonic:math.
Format. .mnm is the project's own format (a 4 KB header, then the tensors), read by the engines in the repo (chat/engine.asm, site/engine.wat). It is not a transformers checkpoint. For llama.cpp and Ollama there are GGUF builds: ollama run supergoatscriptguy/mnemonic gets the 755M, the tags 300m and 126m get the smaller ones, and math gets the math model.
Limitations. The chat models are small: they make confident mistakes (especially with dates and numbers), are weak at math, and lose the thread in long conversations. Each one is better than the one before, but all three share these limits. The math model is the other way around: good at its 38 kinds of problems and no use for anything else.
Made by Supergoatscriptguy.