Denali-AI/README

Denali AI — Vision-Language Models for Garment Classification

Advancing structured attribute extraction from garment images through multi-stage reinforcement learning

---

Abstract

Denali AI develops and benchmarks vision-language models (VLMs) for structured garment attribute extraction — the task of analyzing a garment image and producing a complete JSON object describing 9 key attributes: type, color, pattern, neckline, sleeve length, closure, brand, size, and defect type.

We systematically evaluate the impact of supervised fine-tuning (SFT), Group Relative Policy Optimization (GRPO), and Group-relative Trajectory-based Policy Optimization (GTPO) across multiple model architectures (Qwen3-VL, Qwen3.5-VL, InternVL3, Florence-2) and scales (0.8B to 122B parameters). Our best model, Qwen3-VL-8B SFT+GRPO, achieves 91.3% weighted score with 100% JSON parse rate on the eval_hard_3500 benchmark.

---

Leaderboard

Rank	Model	Architecture	Params	Training	Weighted	SBERT+NLI	JSON%	Throughput
1	Granite4-Vision-3B SFT	Granite4-Vision	4.5B	SFT	101.4%	87.5%	100%	—
2	Qwen3-VL-8B SFT+GRPO	Qwen3-VL	8B	SFT+GRPO	91.3%	78.7%	100%	—
3	Qwen3-VL-2B SFT+GRPO v9	Qwen3-VL	2B	SFT+GRPO	89.5%	78.5%	100%	—
4	Qwen3-VL-8B SFT+GRPO NVFP4	Qwen3-VL	8B	SFT+GRPO+NVFP4	89.5%	77.0%	100%	—
5	Qwen3-VL-8B Instruct (Base)	Qwen3-VL	8B	Zero-shot	87.5%	75.6%	100%	—
6	Qwen3-VL-8B Instruct NVFP4	Qwen3-VL	8B	Zero-shot+NVFP4	87.2%	75.0%	100%	—
7	Qwen3.5-2B Base	Qwen3.5-VL	2B	Zero-shot	84.4%	73.0%	100%	—
8	Qwen3-VL-2B SFT+GRPO v9 NVFP4	Qwen3-VL	2B	SFT+GRPO+NVFP4	84.2%	74.1%	100%	—
9	qwen3.5-0.8b-orr-sft	?	?	?	79.7%	70.5%	100%	—
10	qwen3.5-2b-orr-sft	?	?	?	79.6%	69.9%	100%	—
11	Qwen3-VL-2B Instruct (Base)	Qwen3-VL	2B	Zero-shot	76.4%	66.7%	100%	—
12	InternVL3-2B GRPO+GTPO Full	InternVL3	2B	GRPO+GTPO	72.7%	64.3%	100%	—
13	InternVL3-2B GRPO+GTPO FP8	InternVL3	2B	GRPO+GTPO+FP8	72.2%	63.8%	100%	—
14	InternVL3-2B Base	InternVL3	2B	Zero-shot	71.8%	63.7%	100%	—
15	Moondream2 Base	Moondream	1.6B	Zero-shot	69.8%	61.8%	100%	—
16	Qwen3.5-2B SFT+GRPO+GTPO v8	Qwen3.5-VL	2B	SFT+GRPO+GTPO	65.3%	60.1%	100%	—
17	phi-4-multimodal-sft	?	?	?	65.1%	58.6%	99%	—
18	Qwen3.5-2B SFT v7	Qwen3.5-VL	2B	SFT	63.7%	58.9%	100%	—
19	Qwen3.5-35B GPTQ-Int4	Qwen3.5 MoE	35B (3B)	Zero-shot	50.7%	48.7%	14%	—
20	Qwen3.5-9B NVFP4 v10	Qwen3.5-VL	9B	Zero-shot	47.0%	46.0%	8%	—
21	Qwen3.5-9B SFT NVFP4 v11	Qwen3.5-VL	9B	SFT+NVFP4	46.3%	45.5%	8%	—
22	Qwen3.5-2B NVFP4 v10	Qwen3.5-VL	2B	Zero-shot	42.9%	42.9%	0%	—
23	Qwen3.5-122B-A10B NVFP4	Qwen3.5 MoE	122B (10B)	Zero-shot+NVFP4	42.9%	42.9%	0%	—
24	Qwen3.5-2B SFT NVFP4 v11	Qwen3.5-VL	2B	SFT+NVFP4	42.9%	42.9%	0%	—
25	Qwen3.5-2B SFT+GRPO+GTPO NVFP4	Qwen3.5-VL	2B	SFT+GRPO+GTPO+NVFP4	42.9%	42.9%	0%	—
26	Phi-4 Multimodal NVFP4	Phi-4	5.6B	Zero-shot+NVFP4	42.9%	42.9%	0%	—
27	Qwen3-8B FP8	Qwen3	8B	Zero-shot+FP8	42.9%	42.9%	0%	—
28	granite4-vision-sft-vllm	?	?	?	42.9%	42.9%	0%	—
29	granite4-vision-sft-vllm-deepstack	?	?	?	42.9%	42.9%	0%	—

---

Task Definition

Given a single garment image, the model must extract 9 structured attributes as a valid JSON object:

{
  "type": "t-shirt",
  "color": "navy blue",
  "pattern": "solid",
  "neckline": "crew neck",
  "sleeve_length": "short sleeve",
  "closure": "pullover",
  "brand": "Nike",
  "size": "M",
  "defect_type": "small hole on left shoulder"
}

Field Importance Weights

Not all fields are equally important. The weighted score uses domain-specific multipliers:

Field	Weight	Rationale
Type	2.5x	Critical for inventory routing and categorization
Defect	2.0x	Directly impacts quality control and pricing
Brand	1.5x	Essential for authentication and valuation
Size	1.5x	Required for accurate listing and search
Color, Pattern, Neckline, Sleeve, Closure	1.0x	Standard descriptive attributes

---

Key Results

Per-Field Performance

Accuracy vs Throughput

Key finding: Qwen3-VL-2B v9 achieves the best accuracy-throughput trade-off at 89.5% weighted score and 15.9 samples/s — making it the Pareto-optimal choice for production deployment.

Structured Output Reliability

Fine-tuned models achieve 100% JSON parse rate, while zero-shot baselines (GPTQ, NVFP4) fail to produce valid JSON in 86-100% of cases. This demonstrates that SFT is essential for teaching structured output format, regardless of model scale.

Impact of Training Stages

Left panel: Adding GRPO+GTPO to Qwen3.5-2B improves brand recognition from 15.6% to 24.8% and defect detection from 89.5% to 95.1%, with a +1.6% overall gain.

Right panel: FP8 quantization of InternVL3-2B shows <1% accuracy degradation across all fields while reducing memory footprint, confirming FP8 as a practical deployment optimization.

---

Model Collections

By Architecture

Collection	Models	Description
Qwen3-VL	3	Top-performing Qwen3-VL based models (2B, 8B, 8B-NVFP4)
Qwen3.5-VL	7	Qwen3.5-VL models (0.8B to 122B)
InternVL3	5	InternVL3 models (1B, 2B)
Florence-2	3	Florence-2 encoder-decoder models
Benchmarks	2	Evaluation and training datasets

---

Training Pipeline

All fine-tuned models follow the Denali-AI Multi-Stage RL Pipeline:

                    ┌─────────────────────────────────────────────────┐
                    │           Denali-AI Training Pipeline            │
                    └─────────────────────────────────────────────────┘
                                          │
                    ┌─────────────────────┼─────────────────────┐
                    ▼                     ▼                     ▼
              ┌──────────┐        ┌──────────────┐      ┌──────────────┐
              │  Stage 1  │        │   Stage 2    │      │   Stage 3    │
              │   SFT     │───────▶│    GRPO      │─────▶│    GTPO      │
              │  (LoRA)   │        │  (Rewards)   │      │ (Trajectory) │
              └──────────┘        └──────────────┘      └──────────────┘
                    │                     │                     │
              JSON format          Field accuracy         Coherence &
              acquisition          optimization           regularization

Stage 1: Supervised Fine-Tuning (SFT)

• Method: LoRA (r=16, alpha=32) on frozen base model
• Data: train-10k-balanced-v3 — 10,000 curated samples
• Objective: Teach valid JSON output format and basic field extraction
• Key outcome: 100% JSON parse rate

Stage 2: Group Relative Policy Optimization (GRPO)

• Method: Reward-based RL without a critic model
• Reward engine: 3-layer scoring system
  • Layer 1: JSON validity gate (binary)
  • Layer 2: Structural correctness (20% weight)
  • Layer 3: Per-field content accuracy (80% weight)
• Key outcome: Improved field-level accuracy, especially for challenging fields

Stage 3: Group-relative Trajectory-based Policy Optimization (GTPO)

• Method: Conflict-aware gradient optimization with entropy regularization
• Key outcome: Trajectory-level coherence and reduced field-level conflicts

---

Evaluation Methodology

Benchmark

All models are evaluated on eval_hard_3500 — a curated benchmark of 3,500 challenging garment images selected for diversity in:

• Garment type (tops, bottoms, dresses, outerwear, accessories)
• Visual complexity (patterns, prints, multi-color)
• Edge cases (ambiguous attributes, partially visible labels)

Metrics — Powered by PeakBench

All evaluation is run through PeakBench, our centralized benchmarking platform. Results are automatically synced to HuggingFace model cards via the PeakBench-HF sync bridge. The canonical metric definitions live in peakbench_metrics.json and are shared between both platforms.

Metric	Weight (JSON GT)	Model / Method	Description
Structured Match	60%	Field-level JSON comparison	Per-field presence + value accuracy (null-aware)
SBERT Similarity	25%	all-mpnet-base-v2	Semantic cosine similarity via sentence embeddings
Token Set Ratio	10%	rapidfuzz	Fuzzy word-set overlap (order-independent)
ROUGE-L	5%	LCS F1	Longest common subsequence F-measure
chrF++	—	char+word n-grams	Character and word n-gram F-score
METEOR	—	stems+synonyms	Alignment with stemming and synonym matching
BLEU	—	n-gram precision	BLEU with brevity penalty
Levenshtein	—	edit distance	Normalized character-level edit distance
Hallucination	—	DeBERTa-v3 NLI	Contradiction detection between prompt and response
Consistency	—	SBERT pairwise	Determinism across repeated inference runs

PeakBench Metric Definitions (click to expand)

PeakBench Quality Score

The headline composite metric. For JSON ground truth (our task), weights are: structured match 60%, SBERT similarity 25%, token set ratio 10%, ROUGE-L 5%. Exact case-insensitive match short-circuits to 1.0.

Structured Match

Per-field JSON comparison. Decomposes into field_match_rate (fraction of expected keys present) and value_accuracy (fraction of matched fields with correct values). Null-aware: treats "N/A", "none", "not visible", etc. as equivalent null values.

SBERT Similarity

Semantic cosine similarity using all-mpnet-base-v2 sentence embeddings. Captures meaning-level similarity — "navy blue" and "dark blue" score high despite different strings.

chrF++ Score

Character and word n-gram F-score. Robust for morphologically rich text and partial matches at the character level.

METEOR Score

Alignment-based metric with stemming and synonym matching. Captures paraphrase similarity — "t-shirt" and "tee shirt" score high.

ROUGE-L Score

Longest common subsequence F1. Measures structural word-order overlap between prediction and ground truth.

BLEU Score

N-gram precision with brevity penalty. Standard MT metric, useful as a surface-level quality signal.

Token Set Ratio

Fuzzy word-set overlap via rapidfuzz. Order-independent — "blue navy" matches "navy blue" perfectly.

Levenshtein Ratio

Normalized character-level edit distance. 1 - (edits / max_length). Catches typos and minor variations.

Hallucination Score

NLI contradiction probability between prompt and response using DeBERTa-v3-base-mnli-fever-anli. Higher = more hallucinated. When contradiction > 0.5, the composite score is penalized.

Consistency (Semantic)

Average pairwise SBERT cosine across multiple inference runs on the same prompt. Measures model determinism. 1.0 = perfectly consistent outputs.

JSON Parse Rate

Percentage of outputs that are valid, parseable JSON. Fine-tuned models achieve 100%; zero-shot models often fail at 0-14%.

Throughput

Samples per second via vLLM on NVIDIA RTX PRO 6000 Blackwell (98 GB VRAM), 8 concurrent workers.

Full metric definitions: peakbench_metrics.json

All metrics are computed by PeakBench and automatically synced to HuggingFace model cards. The shared metric config ensures both platforms always display the same numbers.

Evaluation Protocol

• Inference: 8 concurrent workers via OpenAI-compatible API (vLLM)
• Samples: All 3,500 samples, no subsampling
• Compute: NVIDIA RTX PRO 6000 Blackwell (98 GB VRAM)
• Reproducibility: Fixed prompts, deterministic sampling (temperature=0)

---

Key Findings

1. Best model: Granite4-Vision-3B SFT achieves 101.4% weighted score with 100% JSON parse rate on 3,500 hard samples.

2. Granite4-Vision dominates. Best Granite4-Vision (101.4%) leads best Qwen3-VL (91.3%) by 10.1pp. Architecture rankings: Granite4-Vision (101%), Qwen3-VL (91%), Qwen3.5-VL (84%), ? (80%), InternVL3 (73%).

3. SFT is essential for structured output. Fine-tuned models: 76% avg JSON parse rate, best 101.4%. Zero-shot models: 52% avg JSON parse, best 87.5%. Training adds +13.9pp at the top.

4. NVFP4 quantization costs 5.3pp on average (max 5.3pp) across 1 model pairs, while reducing size ~60% and increasing throughput ~50%.

5. Hardest fields (on best model): neckline (79%), pattern (80%), type (80%). Easiest: brand (96%), defect (97%), size (100%).

6. Scale vs efficiency. Best large (Qwen3-VL-8B SFT+GRPO: 91.3%) beats best small (Qwen3-VL-2B SFT+GRPO v9: 89.5%) by 1.8pp — small model is highly competitive for edge deployment.

7. Benchmark coverage: 29 models across 9 architectures, 12 fine-tuned + 17 zero-shot/quantized.

---

Research Directions & Future Work

Near-Term Improvements

Direction	Expected Impact	Rationale
SFT+GRPO on Moondream	+5-15pp	Zero-shot at 69.8%, fine-tuning consistently adds significant gains
SFT+GRPO on Qwen3.5 MoE	+5-15pp	Zero-shot at 50.7%, fine-tuning consistently adds significant gains
SFT+GRPO on Phi-4	+5-15pp	Zero-shot at 42.9%, fine-tuning consistently adds significant gains
SFT+GRPO on Qwen3	+5-15pp	Zero-shot at 42.9%, fine-tuning consistently adds significant gains
NVFP4 quantize Granite4-Vision-3B SFT	-1-2pp, +50% speed	At 101.4%, no quantized variant exists yet
NVFP4 quantize Qwen3.5-2B SFT+GRPO+GTPO v8	-1-2pp, +50% speed	At 65.3%, no quantized variant exists yet
NVFP4 quantize Qwen3.5-2B SFT v7	-1-2pp, +50% speed	At 63.7%, no quantized variant exists yet
GTPO on Qwen3-VL-8B SFT+GRPO	+1-3pp	Currently SFT+GRPO only, GTPO adds trajectory coherence
GTPO on Qwen3-VL-2B SFT+GRPO v9	+1-3pp	Currently SFT+GRPO only, GTPO adds trajectory coherence

Architecture Exploration

Models not yet benchmarked — recommended based on current findings:

Model	Parameters	Why Promising
Qwen3-VL-3B-Instruct	3B	Same family as our #1 (Granite4-Vision), mid-range scale
InternVL3-8B	8B	Larger InternVL — may close gap to Qwen3-VL at same scale
InternVL3-4B	4B	Mid-range InternVL — potential efficiency sweet spot
SmolVLM2-2.2B-Instruct	2.2B	HuggingFace's efficient VLM — strong structured output
PaliGemma2-3B	3B	Google VLM with excellent OCR — may improve brand/size fields
Phi-4-multimodal-instruct	5.6B	Microsoft VLM — needs SFT (zero-shot JSON fails)
MiniCPM-V-2.6	2.8B	Strong small VLM with good OCR capabilities
Molmo-7B-D	7B	Allen AI VLM — strong visual grounding, may help with defect detection
Idefics3-8B	8B	HuggingFace VLM — instruction-following optimized
DeepSeek-VL2-Small	3B	DeepSeek's latest compact VLM — strong reasoning

Long-Term Research

1. Ensemble routing: Route each field to its best-performing model architecture
2. Multi-image input: Front + back + tag images simultaneously for higher brand/size accuracy
3. Curriculum learning: Progressive difficulty — easy garments first, hard edge cases last
4. Synthetic data: Use 122B models to generate training labels at scale
5. Active learning: Prioritize annotation of samples where models disagree most
6. Guided JSON decoding: Constrained generation to force valid JSON without training

Key Open Questions

• Why does Granite4-Vision outperform Qwen3-VL by 10.1pp at similar scale? Vision encoder, cross-attention, or training data?
• Can RL gains (GRPO/GTPO) be amplified beyond current levels with better reward engineering?
• Is there a parameter sweet spot between 2B and 8B where accuracy saturates?
• Would domain-specific pre-training (garment images) outperform general VLM fine-tuning?
• closure averages only 52% across top-5 models — is the ground truth noisy, or is this genuinely hard?

---

Datasets

Dataset	Samples	Purpose	Link
eval_hard_3500	3,500	Evaluation benchmark (hard subset)	Link
train_10k_balanced_v3	10,000	Training data (balanced sampling)	Link

---

Last updated: 2026-04-04 03:59 UTC

---

Citation

@misc{denali-ai-2026,
  title={Structured Garment Attribute Extraction via Multi-Stage Reinforcement Learning},
  author={Denali AI},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/Denali-AI}
}

License

All models and datasets are released under the Apache 2.0 License.

Contact

• Organization: Denali Advanced Integration
• Issues: GitHub
• HuggingFace: Denali-AI