CoolFace
Modelpublic

CodeSoft/sorbet-v2-25m

sourceHugging Faceapache-2.0updated 27d agoView on Hugging Face
1likes1.7kdownloads
Model Card

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/645aad59c4acfcf664022df5/7_BzxM3bmTilI1xRzElnB.png" alt="Sorbet V2 Header" width="700">

<h1>Sorbet v2 25M</h1>

<p> 25M-parameter Qwen2-style decoder LM, warm-started from Sorbet-25M and continued-trained in two runs. </p> </div>

Architecture graph

<a href="https://hfviewer.com/CodeSoft/sorbet-v2-25m?utmsource=huggingface&amp;utmmedium=embeddedmodelcard&amp;utmcampaign=CodeSoftsorbet-v2-25mcard" target="blank" rel="noopener"> <img src="https://hfviewer.com/api/card.svg?source=CodeSoft%2Fsorbet-v2-25m&amp;granularity=0" alt="Architecture graph for CodeSoft/sorbet-v2-25m. Open in hfviewer" width="100%" /> </a>

Architecture

Identical to Sorbet-25M: stock Qwen2 throughout, no custom code paths, natively supported by both transformers and llama.cpp.

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

v2 continues the v1 checkpoint through two training runs:

LegData mix (tokens)LR schedule
cpt2fineweb-edu 70% / infiwebmath 10% / DCLM-baseline 20%, 0.8B tokcosine, 8-bit AdamW
v2-finalFineWeb-HQ 65% / DCLM-baseline 20% / FineMath-4+ 15%, 1.7B tokcosine peak 1e-4, fp32 AdamW

Block-shuffled at 131,072 tok/step.

Benchmarks

All numbers zero-shot via lm-evaluation-harness, bf16, identical settings across checkpoints.

TasknRandomaccacc_norm
HellaSwag10,04225%26.55 ±0.4426.63 ±0.44
ARC-easy2,376~25%30.30 ±0.9429.92 ±0.94
ARC-challenge1,172~25%18.60 ±1.1422.44 ±1.22
PIQA1,83850%54.52 ±1.1653.32 ±1.16
ArithMark-3.01,00025%32.90 ±1.4833.00 ±1.49

Notes:

  • —Every score is at or above the Sorbet-25M baseline within error bars.
  • —ArithMark-3.0 (AxiomicLabs/Arithmark-3.0) remains the strongest relative result (+8 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-v2-25m"
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.