math-think-o7b-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 | 51 | 124 |
| Same-run Think Base | 34 |
Think mode; stop ids [100257, 100265]; O7B eval uses OLMO3-VLLM-PATCH-v1 / vLLM 0.27.1. Same-run Think Base is think_vllm027 34/240 (first think Base attempt failed).
θ_0 is allenai/Olmo-3-1025-7B rev 996971efdc504b81f0a6caf73a6c92f976254b9c.
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 [100257].
Identity
| Field | Value |
|---|---|
| Arm | O7B-THINK-A3B |
| Updates / tokens | 987 / 64,632,872 |
Merged model.safetensors sha256 | 9f75574aed1be033f277a0384b97f4edef3f8ac85401c6781e59c1d46737d526 |
| Endpoint adapter sha256 | 9da5899df6a1d6d7cfb9e412085eed9a85693bd0fd848645b176e34deed0d8e2 |
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-o7b-20260908", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("modrill/math-think-o7b-20260908")