math-think-q4b-20260908
Public freeze of Math Think 2ep endpoint for ICLR 2027 task-vector work. Not a chatbot. Endpoint is the score. Do not promote 1ep / 1.5ep milestones.
run_id=math_six_arms_train_v3_eot_20260908. Sibling of modrill/nothink-src-*-20260908; does not overwrite those repos.
Score (Exact-240)
AIME24+25 × seeds 42–45, EvalScope reviews, n=240. Authority field evalscope_reviews.correct.
| Model | Official /240 | cap |
|---|---|---|
| This 2ep endpoint | 40 | 169 |
| Same-run Think Base | 21 |
Think mode; stop ids [151643, 151645]; vLLM 0.25.1 audited runtime.
θ_0 is Qwen/Qwen3-4B-Base rev 906bfd4b4dc7f14ee4320094d8b41684abff8539.
Recipe
OpenR1 11750 rows × 2ep (unique problems 5875), LoRA r64/α128, lr 1e-4, TPU 65536, seed 42, EOT (Qwen tail 151643 / O7B tail 100257), B-rows both sides [151643, 151667, 151668].
Identity
| Field | Value |
|---|---|
| Arm | Q4B-THINK-A3B |
| Updates / tokens | 977 / 63,999,854 |
Merged model.safetensors sha256 | c83a531ed08c84ee6999dbc75cc8cf5e362676f317328889e9a731b56f434552 |
| Endpoint adapter sha256 | 3723bbe789bf7b07edadb84dddab2de9b1a13652d3c0778284c358608522e2d6 |
Root of merged weights is this repo. Endpoint LoRA is in adapter/. MERGE_RECEIPT.json is the merge audit. trainer_state is not uploaded.
Load
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("modrill/math-think-q4b-20260908", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("modrill/math-think-q4b-20260908")