Skip to content

[P2] CoT baseline + perturbation harness #16

Description

@duckyquang

Phase: P2
Owner(s): Akash


Goal: Token-level CoT comparison baseline so paper claims are fair.

Where: marc/eval/baselines/cot_baseline.py, marc/eval/baselines/perturb.py.

Parallel track: Day 1 of P2. Use same test split JSON as Sparsh — no need to wait for Stage-B model for baseline runs.

  1. Implement CoT baseline wrapper (API or local model) matching same problem set.

  2. Implement constant perturbation: change coefficients in G, re-run solver, log success rate drop.

  3. Export results to results/p2_main/cot_baseline.json for Sparsh to merge into figures.

  4. Document model + prompt used in marc/eval/baselines/README.md.

Done when: CoT baseline JSON exists and Sparsh has merged it into summary table.

Push to: PR akash/p2-baselinemain.


Auto-created from the MARC task board (index.html). Update the board if scope changes.

Metadata

Metadata

Assignees

Labels

P2 · checker RLPhase 2 — checker fine-tuning / GRPOparallel-trackCan start in parallel — no blocking wait

Type

No type

Fields

No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions