Field	Value
Base model	`Qwen/Qwen2.5-VL-3B-Instruct`
Adapter type	PEFT LoRA
Agent mode	`no_turn`
SFT mode	`no_turn`
Ablation protocol	`from_base` (from-base learning curve)
Eval protocol	`canonical_clean_v1`
N (train examples)	200
TrackIO run	`20260809_154043_fa7dba37`
Generated (UTC)	2026-08-09T19:36:08.332971+00:00

Training

Field	Value
Dataset	`abcasas/VIGIA-QA-dataset` (entity-sanitized gold; no `<voice_>` / `<face_>`)
Split	train 90% (eval holdout never used as reward)
Gold column	`gold_sft_no_turn`
Epochs	2
Learning rate	5e-05
LoRA r / alpha	16 / 32
Hardware preset	`rtx5060`
Eval frames (N-curve)	`timeline_16`

Evaluation (post_sft)

Canonical clean: M_full QA, A/B-balanced eval view, stratified 7 scenes, judge ON.

Metric	Value
n_eval	20
balanced_qa	0.7000
pct_fmt_tags	1.0000
epi_qa_acc	0.7500
sem_qa_acc	0.6500

Output format

Final turn must use [Answer] with structured tags:

<think>...</think>
[Answer]
<episodic_memory>...</episodic_memory>
<semantic_memory>...</semantic_memory>

Inference (minimal)

python

import torch
from peft import PeftModel
from transformers import AutoProcessor, BitsAndBytesConfig, AutoModelForVision2Seq

base_id = "Qwen/Qwen2.5-VL-3B-Instruct"
adapter_id = "abcasas/vigia-sft-no-turn-n200"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)
model = AutoModelForVision2Seq.from_pretrained(
    base_id, device_map="auto", quantization_config=bnb, torch_dtype=torch.bfloat16
)
model = PeftModel.from_pretrained(model, adapter_id)
processor = AutoProcessor.from_pretrained(
    base_id, min_pixels=256*256, max_pixels=448*448, use_fast=True
)
processor.tokenizer.padding_side = "left"

# Prefer the package control loop for no_turn (16 frames, max_tool_turns=0):
# from vigia_sft_no_turn import ExperimentConfig
# from vigia_sft_no_turn.control_loop import run_control_loop
# traj, resp, _ = run_control_loop(example, model, processor, agent_mode="no_turn", max_tool_turns=0, ...)

Limitations

Adapter requires the listed base model weights.
Does not include Bedrock QA judge.
Trained on VIGIA-QA surveillance clips; not a general VQA model.
Do not confuse with GRPO repos (vigia-grpo-no-turn-*).
Metrics above are from protocol canonical_clean_v1 (clean gold). Older Aug-07 ladder cards used leaky gold and are not comparable.

Citation

Placeholder for thesis / article — update before publication.

Canonical clean campaign

protocol_id: canonical_clean_v1
gold: entity-sanitized (no <voice_*> / <face_*>)
eval: A/B 50/50 view, M_full, F16, judge ON, max_tool_turns=0
TrackIO project: vigia-sft-no-turn-clean
Local curve: outputs/sft_no_turn_ladder_clean/
REPORT: outputs/reeval_canonical_clean/REPORT.json