purefall/shotpath-boundary-cot-v8-existing-full-core-ablation-data-20260728

ShotPath Boundary CoT v8 Existing Full-Core Ablation

Status: training candidate

This experiment tests whether explicit visible reasoning improves the canonical open photography diagnosis task when no new teacher labels are added.

Task

Both variants use the canonical full-review prompt from D:\data\shotpath\SHOTPATH_TASK_AND_DATA_CONTRACT.md. Directed rows add only the allowed dimension scope line.

Data

• 1,532 directed exposure rows drawn from existing strong supervision.
• 252 unique complete five-dimension rows, repeated three times.
• 2,288 total exposure rows.
• 821 unique images.
• Zero overlap with canonical Eval100 image hashes.
• No new GPT-5.4 or other teacher calls.

The complete five-dimension core contains only rows whose five dimensions were retained by the independent visual audit, contain no uncertain decision, and have at least two issue dimensions. Tiers A and B are defined in existing_five_dimension_reuse_audit_20260728.

Controlled Ablation

train_cot.jsonl uses a compact fixed target:

Dimension
Evidence
Assessment
Decision
Problem/Action when supported

Final:
{"items":[...]}

train_no_cot.jsonl uses only the identical final items JSON. Row IDs, images, prompts, ordering, and final labels are identical between variants.

Sampling

Full-review rows account for 33.04% of exposure rows. This intentionally exceeds the historical v5 directed-main ceiling because the experiment targets five-dimension reasoning serialization. The source and issue-dimension exposure mixtures remain within the canonical contract tolerances.

Training

Train Qwen3-VL-8B-Instruct LoRA with ordinary next-token cross entropy. The two variants must use identical hyperparameters and seed. Evaluate with deterministic generation and the canonical Eval100 runner. Parse only model raw output and never use GT fallback.