Modelpublic

modrill/math-nothink-q4b-20260908

sourceHugging Faceapache-2.0updated 15d agoView on Hugging Face

0likes203downloads

Model Card

math-nothink-q4b-20260908

Public freeze of Math NoThink source expert θ_s for ICLR 2027 task-vector transfer. Not a chatbot. Endpoint is the score. Do not promote milestones.

run_id=math_six_arms_train_v3_eot_20260908. Replaces the private 20260902 src freeze for this v3 EOT recipe; does not overwrite nothink-src-*-20260902.

Score (Exact-240)

AIME24+25 × seeds 42–45, EvalScope reviews, n=240.

Model	Official /240
This endpoint	49
Same-run Base	22

τs = θs − θ0. θ0 is Qwen/Qwen3-4B-Base rev 906bfd4b4dc7f14ee4320094d8b41684abff8539.

Identity

Field	Value
Arm	`Q4B-NOTHINK-EP-X`
Updates / tokens	165 / 10,788,766
Recipe	LoRA r64/α128, TPU 65536, 2ep row-matched, seed 42, Qwen tail `151643` / O7B tail `100257`, B-rows both sides [151643, 151667, 151668]
Merged `model.safetensors` sha256	`f58598701c77f8b82d8ac31abf35689c1cf097dde8d9652bf446a8fa4b2e1465`
Endpoint adapter sha256	`b6e1a5fe40eb2407880953e746ce4788af62034ee395918c4fee179c8770bf54`

Root of merged weights is this repo. Endpoint LoRA is in adapter/. MERGE_RECEIPT.json is the merge audit.

Load

python

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("modrill/math-nothink-q4b-20260908", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("modrill/math-nothink-q4b-20260908")