Downloads · 30 days
11
12% of all-time downloads
vigia-sft-no-turn-n600
abcasas/vigia-sft-no-turn-n600
vigia-sft-no-turn-n600 is a machine learning model from abcasas. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft.
LoRA adapter for VIGIA SFT noturn (line vigiasftnoturn).
All-time downloads
94
Public
Repo size
1.6 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt240 MB · 47%
At a glance
- Library
- peft
- Access
- Public
- Created
- Jul 9, 2026
- Updated
- Aug 9, 2026
- SHA
- 627678f6
Base models
Model card
From the Hugging Face model README
vigia-sft-no-turn-n600
LoRA adapter for VIGIA SFT no_turn (line vigia_sft_no_turn).
Identity
| 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) | 600 |
| TrackIO run | 20260809_154043_fa7dba37 |
| Generated (UTC) | 2026-08-09T19:36:09.854519+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 | 0.6500 |
| pct_fmt_tags | 1.0000 |
| epi_qa_acc | 0.6500 |
| 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)
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-n600"
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
- Library
- peft
- Languages
- en
- Created
- Jul 9, 2026
- Updated
- Aug 9, 2026