CoolFace
Modelpublic

CodeSoft/sorbet-25m

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
4likes1.2kdownloads
Model Card

Sorbet-25M

Architecture graph

<a href="https://hfviewer.com/CodeSoft/sorbet-25m?utmsource=huggingface&amp;utmmedium=embeddedmodelcard&amp;utmcampaign=CodeSoftsorbet-25mcard" target="blank" rel="noopener"> <img src="https://hfviewer.com/api/card.svg?source=CodeSoft%2Fsorbet-25m&amp;granularity=0" alt="Architecture graph for CodeSoft/sorbet-25m. Open in hfviewer" width="100%" /> </a>

From-scratch ~25M-parameter Qwen2-style decoder LM trained in under 4h on a single RTX 5060 Ti (16GB).

Architecture

Params25,185,920 (~87% non-embedding)
Layers / hidden14 / 384
AttentionGQA 6 heads / 2 KV heads, RoPE θ=100k
FFN1024 (SwiGLU)
Context4096
Vocab8,192 custom byte-level BPE (tied embeddings)
Precisionbf16

Training data

0.8B-token weighted mix: fineweb-edu 70% / infiwebmath 10% / DCLM-baseline 20%, block-shuffled. ~3000 steps at 262,144 tok/step, cosine LR, 8-bit AdamW.

Benchmarks

TasknRandomaccacc_norm
HellaSwag10,04225%26.52 ±0.4426.12 ±0.44
ARC-easy2,376~25%29.50 ±0.9429.59 ±0.94
ARC-challenge1,172~25%17.66 ±1.1122.95 ±1.23
PIQA1,83850%54.46 ±1.1653.43 ±1.16
ArithMark-3.01,00025%32.70 ±1.4832.90 ±1.48

Notes:

  • —ArithMark-3.0 (AxiomicLabs/Arithmark-3.0) is the strongest relative result (+7.9 pts over random), consistent with the math share of the pretraining mix.
  • —ARC-challenge raw accuracy sits below chance due to a length bias in unnormalized scores; acc_norm is the meaningful metric there.

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "CodeSoft/sorbet-25m"
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16").to("cuda")
tok = AutoTokenizer.from_pretrained(repo, subfolder="tokenizer")
ids = tok("Once upon a time", return_tensors="pt").input_ids.cuda()
print(tok.decode(model.generate(ids, max_new_tokens=64)[0]))

Limitations

Expect shallow world knowledge and weak performance on knowledge-heavy benchmarks due to the model's small parameter count and limited training budget.

License

Apache-2.0.