Skip to content

baseline_local

diapason.agents.hybrid.baseline_local

BaselineLocalAgent — local-only reference for the hybrid ablation.

Mirror of :class:BaselineCloudAgent (baseline_cloud.py) but the entire trajectory runs on the local vLLM model. No cloud teacher / router / advisor is involved — this is the "what does the local model do by itself?" floor in the n=100 ablation matrix.

On GAIA the agent makes one local call with the formatted prompt (which already carries the FINAL ANSWER: reminder from _prompts.format_gaia) and returns the text. On SWE-bench-Verified the agent delegates to :func:run_swe_agent_loop with backbone="local" so the model gets to run bash and read the repo — same wiring as the mini-swe-agent cells but driven by the local model.

Construction args mirror :class:LocalCloudAgent. The local block in the cell registry determines the local model + endpoint; cloud is accepted for schema compatibility but unused (and cost_usd is always 0 — local inference is free).

BaselineLocalAgent

Bases: LocalCloudAgent

Local-only baseline. cloud_* fields are ignored.

Configurable knobs via cfg:

  • local_max_tokens (int, default 4096): max_tokens per GAIA call and per turn of the SWE agent loop.
  • local_temperature (float, default 0.0): sampling temperature for the local model.
  • swe_use_agent_loop (bool, default True for SWE): if False the SWE branch falls back to a one-shot blind patch (not recommended; kept for parity with other agents).
  • swe_max_turns (int, default 30): SWE-bench loop turn cap.
  • swe_bash_timeout_s (int, default 120): bash timeout per turn.
  • swe_turn_max_tokens (int, default 4096): max_tokens per agent turn inside the SWE loop. Falls back to local_max_tokens.