CodeSoft/sorbet-v2-25m

Sorbet v2 25M

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

Architecture graph

Architecture

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

Params	25,185,920 (~87% non-embedding)
Layers / hidden	14 / 384
Attention	GQA 6 heads / 2 KV heads, RoPE θ=100k
FFN	1024 (SwiGLU)
Context	4096
Vocab	8,192 custom byte-level BPE (tied embeddings)
Precision	bf16

Training

v2 continues the v1 checkpoint through two training runs:

Leg	Data mix (tokens)	LR schedule
cpt2	fineweb-edu 70% / infiwebmath 10% / DCLM-baseline 20%, 0.8B tok	cosine, 8-bit AdamW
v2-final	FineWeb-HQ 65% / DCLM-baseline 20% / FineMath-4+ 15%, 1.7B tok	cosine 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.

Task	n	Random	acc	acc_norm
HellaSwag	10,042	25%	26.55 ±0.44	26.63 ±0.44
ARC-easy	2,376	~25%	30.30 ±0.94	29.92 ±0.94
ARC-challenge	1,172	~25%	18.60 ±1.14	22.44 ±1.22
PIQA	1,838	50%	54.52 ±1.16	53.32 ±1.16
ArithMark-3.0	1,000	25%	32.90 ±1.48	33.00 ±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

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.