purefall/shotpilot-e5-full-evidence-qwen3vl8b-lora

ShotPilot Full-Five-Evidence Qwen3-VL-8B LoRA

This archive contains the three epoch checkpoints and reproducibility artifacts for the strict full-five-evidence ablation.

Training

• 2,483 training rows
• Qwen3-VL-8B-Instruct
• LoRA rank 64, alpha 32, dropout 0.05
• 3 epochs, checkpoints 156/312/468
• learning rate 5e-6, weight decay 0.1, warmup ratio 0.03, cosine schedule
• global batch size 16, BF16, max sequence length 4096, image size 672, seed 42
• assistant target: all five canonical Evidence observations followed by aggregated Problem and Action

Evaluation

Each checkpoint was evaluated on the fixed Eval100_new set with the fixed free-review prompt, deterministic decoding, and max_new_tokens=768. Ground-truth fallback and pair judging were disabled.

Primary outputs are preserved unchanged. Separate repaired outputs replace only rows with loop_8gram_max >= 4, using deterministic escalating repetition controls:

• round 1: no-repeat-ngram 8, repetition penalty 1.05
• round 2 when needed: no-repeat-ngram 6, repetition penalty 1.10
• round 3 when needed: no-repeat-ngram 4, repetition penalty 1.15

Loop audit:

Checkpoint	Primary loops	Repaired loops	Repaired rows
156	36	0	36
312	12	0	12
468	5	0	5

The repaired files are decoding interventions and must not be reported as the untouched primary model output. Every primary and repaired file contains exactly 100 unique evaluation IDs.

Archive Contents

• checkpoints/: adapter weights, adapter config, trainer state, and checkpoint README for 156/312/468
• nonweights.tar: training/evaluation scripts, train and evaluation JSONL manifests, summaries, logs, and token audit
• manifest.json: local/remote byte sizes and SHA256 checksums

The archive excludes optimizer state, base-model weights, and images.