deepgraph.toml

# DeepGraph unified configuration.
#
# Precedence: process environment / .env overrides this file.
# Keep real API keys and SSH passwords in .env or secret storage; leave the
# secret fields here empty unless you are using a private local checkout.

[profile]
active = "machine_learning"

[app]
name = "DeepGraph"
subtitle = ""
root_node_id = "ml"

[paths]
db_path = "{project_root}/deepgraph.db"
workspace_dir = "{project_root}/workspace"
pdf_cache_dir = "{project_root}/workspace/pdfs"
experiment_workdir = "{home}/sciforge_runs"
idea_workspace_dir = "{home}/deepgraph_ideas"
manuscript_workdir = "{home}/deepgraph_manuscripts"
reference_pdf_corpus_dir = "{project_root}/../../workspace/pdfs"
iclr2026_template_dir = "{project_root}/third_party/iclr2026/iclr2026"

[database]
url = ""
pgvector_embedding_dim = 1536

[runtime]
python = ""

[tracking]
mlflow_uri = ""

[llm]
use_tabcode = false
protocol = "responses"
base_url = "https://api2.tabcode.cc/openai"
api_key = ""
model = "gpt-5.4"
rpm = 0
extra_providers_json = ""
reasoning_effort = "medium"
connect_timeout_seconds = 30
request_timeout_seconds = 300
transient_retries = 2
transient_backoff_seconds = 5
transient_cooldown_seconds = 180
max_input_tokens = 900000
max_output_tokens = 32000

[llm.secondary]
enabled = false
protocol = "responses"
base_url = ""
api_key = ""
model = "gpt-5.4"
rpm = 0

[minimax]
api_key = ""
base_url = "https://api.minimaxi.com/v1"
model = "MiniMax-M2.7-highspeed"
rpm = 18

[prompts]
version = "insight_v1"

[arxiv]
categories = ["cs.AI", "cs.LG", "cs.CL", "cs.CV", "cs.MA", "stat.ML"]
max_results_per_query = 100

[grobid]
base_url = "http://127.0.0.1:8070"
request_timeout_seconds = 300

[pdf]
text_backend = "auto"

[pipeline]
concurrency = 30
sleep_between_papers = 1
incremental_insight_every = 20
max_retryable_failures = 12
event_poll_seconds = 5

[graph]
grounding_min_store_score = 0.0
contradiction_requires_grounded = false
backfill_on_start = true
refresh_merge_candidates_on_start = true
paper_cluster_min_papers = 10

[discovery]
tier1_candidates = 5
tier2_problems = 8
tier2_papers = 5
min_tier2_backlog = 3
auto_trigger_papers = 200
bulk_tier1_candidates = 12
bulk_tier1_overlaps = 80
bulk_tier1_patterns = 60
bulk_tier2_problems = 15
bulk_tier2_plateaus = 35
bulk_tier2_limit_nodes = 30

[idea]
evidence_min_non_numeric = 2

[experiment]
time_budget_seconds = 300
max_iterations = 100
reproduction_iters = 3
proxy_data_fraction = 0.15
proxy_max_epochs = 10
early_stop_threshold = 0.20
refute_min_iters = 30
plateau_patience = 12
require_real_benchmark = true
allow_synthetic_fallback = false
real_llm_model = "Qwen/Qwen2.5-7B-Instruct"
real_benchmark_dataset = "openai/gsm8k"
real_benchmark_dataset_config = "main"
real_benchmark_max_examples = 64
real_benchmark_seeds = 3
real_benchmark_time_budget_seconds = 3600
validation_benchmark_max_examples = 64
validation_benchmark_seeds = 3
validation_benchmark_methods = "Vanilla Direct Answering,Always-Reason Chain-of-Thought,CGGR,CGGR/no_counterfactual_delta"
reproduction_repair_max_rounds = 8
allow_smoke_validation = false
evosci_verify_timeout = 900
require_evoscientist = false

[codex]
exec_enabled = true
cli_path = ""
model = ""
timeout_seconds = 900

[gpu]
mode = "single_host"
worker_slots = 4
visible_devices = []
default_model = "H20"
default_vram_gb = 96
poll_seconds = 10
stale_recovery_poll_seconds = 120
job_timeout_seconds = 14400

[gpu.remote]
ssh_host = ""
ssh_port = 22
ssh_user = ""
ssh_password = ""
base_dir = "/root/deepgraph-remote-worker"
python = "python"
auto_pip_install = true
setup_timeout_seconds = 3600

[manuscript]
latex_template = "iclr2026"
submission_bundle_formats = ["conference"]
backend = "paper_orchestra"
iclr2026_template_files = [
  "iclr2026_conference.sty",
  "iclr2026_conference.bst",
  "math_commands.tex",
  "natbib.sty",
  "fancyhdr.sty",
]

[paper_orchestra]
semantic_scholar_api_key = ""
paperbanana_cmd = ""
refinement_iters = 4

[auto_research]
pipeline_enabled = false
pipeline_batch_size = 100
pipeline_interval_seconds = 120
pipeline_start_delay_seconds = 10
enabled = true
interval_seconds = 300
max_active = 1

[web]
host = "0.0.0.0"
port = 8080