text stringclasses 1
value |
|---|
mirror-ok |
Octopus: Training Environments, Harbor Bundles & Deep-Search BM25 Corpora
🌐 Project Page · 📄 Technical Report (PDF) · 💻 GitHub Code & Runtime
Official training queries, pre-built BM25 retrieval indices, containerized Harbor SWE task bundles, and frozen runtime overlays for Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning (Tianming Sha, Junnan Li, Siqi Zhao).
Overview
Octopus is a co-designed serving runtime, GPU memory manager, and reinforcement learning (RL) framework (SGLang + AReaL) that executes long-horizon multi-agent reasoning as a stateful, GPU-resident execution DAG over immutable radix KV cursors (fork, rewind, read_thread, wait, done):
- Server-Resident Radix Cursors (
89.6%GPU KV Hit Rate): Branches fork inO(1)zero copy with isolated32Kgrowth budgets and restore exact hybrid recurrent checkpoints (h_k), cutting wire traffic by 98.3% and accelerating128Kprompt assembly by 249× (358.5 → 1.44 ms). - Hijack-Free Zero-Copy KV Mounting (
92.2%Info Retention,21×Lower Hijack): Splices sibling KV pages exclusively into full-attention layers under monotonic position offsets while recurrent layers (Gated DeltaNet/Mamba-2) skip mounted spans—allocating 0 extra KV pages and cutting off-task hijacking from 57.9% to 2.7%. - Prefix-Tree Packed
DAG-GRPO&4BFrontier Results: Deduplicates shared prefixes via prefix-deduplicating loss masks while keeping mounted KV edges differentiable (32×lower rollout–training KL at4.45 × 10⁻⁴). Trained on Qwen3.5-4B under sparse BM25 and pure0/1outcome rewards, Octopus reaches 49.6% on BrowseComp-Plus (+17.5 pts over143Ksingle-agent GRPO) and 51.2% on OpenResearcher, outperforming dense-retrieval browser agents up to 250× larger.
Repository Branches (--revision)
This dataset repository organizes Octopus artifacts across two dedicated Git branches (search and code) so users only pull the footprint required for their workload:
Branch (--revision) |
Approx. Size | Contents |
|---|---|---|
search |
~57 GB | OpenResearcher training & validation query splits (3,040 train / 500 val) + pre-built sparse BM25 indices and document stores for OpenResearcher-Corpus (14.9M docs) and BrowseComp-Plus (100,195 docs, 830 test queries). |
code |
~330 MB | Containerized Harbor SWE task directories (bundles-v2, coupled-a, cooperbench-v2) and env-bundle (frozen dependency wheels & SGLang 0.5.19 / AReaL cursor-mount overlays). |
Quickstart & Download
Option 1: Automated Setup via octopus Repo (Recommended)
git clone https://github.com/shatianming5/octopus.git && cd octopus
# Pull code/environment bundle (--search also pulls the 57 GB BM25 search corpus & training queries)
bash scripts/deploy.sh $ROOT fetch --search
bash scripts/deploy.sh $ROOT install
bash scripts/deploy.sh $ROOT check --gpu 0
Option 2: Direct Download via huggingface-cli
# Optional: if huggingface.co connections reset on your GPU cluster, configure a mirror first:
# export HF_ENDPOINT=https://hf-mirror.com
# 1) Pull Harbor SWE task bundles & frozen runtime env-bundle (~330 MB)
huggingface-cli download Tommysha/octopus-train \
--repo-type dataset \
--revision code \
--local-dir $ROOT/code-data
# 2) Pull OpenResearcher training queries & pre-built BM25 indices (~57 GB)
huggingface-cli download Tommysha/octopus-train \
--repo-type dataset \
--revision search \
--local-dir $ROOT/search-data
Note on Containerized Coding Tasks: Container images are not stored inside this repository. For coding/SWE tasks (
codebranch), pull the task-specifieddocker_imagetargets (udocker/Singularityrootless supported) and runprebuild_goldens.pyas documented in thecodebranch README andshatianming5/octopus.
Citation
@techreport{sha2026octopus,
title = {Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning},
author = {Sha, Tianming and Li, Junnan and Zhao, Siqi},
institution = {Stony Brook University and University of Wisconsin--Madison and University of Minnesota},
year = {2026},
month = nov,
url = {https://shatianming5.github.io/octopus/},
pdf = {https://shatianming5.github.io/octopus/assets/Octopus.pdf}
}
- Downloads last month
- 233