Orqo Hype v1

Orqo Hype v1

Orqo Hype is a specialist composition model for the HyperFrames stack. Given a creative brief, it produces a staged pipeline:

design JSON → storyboard → scene HTML / GSAP

suitable for HyperFrames assemble → lint → render (MP4).

Product	Orqo Hype
Version	v1
HF id	cristianoaredes/orqo-hype-v1
Visibility	Public
Internal train id	sft-v9b-mix
Load base (compatibility only)	Qwen/Qwen2.5-Coder-7B-Instruct

Branding: the product name is Orqo Hype. The foundation model is listed only for PEFT / merge load compatibility — it is not part of the product name.

---

Verified repository contents

Live Hub inventory verified on 2026-07-17:

Path	Format	Intended use
lora/	PEFT adapter	Load with the compatible base model
merged/	Merged Hugging Face weights (4 safetensors shards)	Local Transformers-compatible inference without PEFT
gguf/	Q4_K_M GGUF	Local llama.cpp-compatible inference
mlx-4bit/	MLX 4-bit, group size 64, 4.501 bits/weight	Local Apple Silicon inference with mlx-lm
coreml/	Core ML 4-bit palettized .mlpackage	iOS 18+/macOS 15+ compatibility research; smoke-tested on macOS only

orqo-hype-v1/
├── README.md                 ← this card
├── lora/                     # PEFT adapter + tokenizer + train_metrics
├── merged/                   # full HF weights
├── gguf/                     # Q4_K_M GGUF + metadata
├── mlx-4bit/                 # MLX 4-bit weights + export metadata
└── coreml/                   # Core ML package + export metadata

Evidence: live Hugging Face Hub inventory plus mlx-4bit/export_metadata.json and coreml/export_metadata.json in this repository. The 2026-07-12 packaging record covers lora/, merged/, and gguf/; the live 2026-07-17 inventory additionally verifies MLX and Core ML.

---

Local usage

Cost boundary: this card does not advertise or invoke a hosted inference endpoint. The examples below download the public model artifacts and run them on hardware controlled by the user. They do not call a project-owned paid inference service.

1) Load LoRA (PEFT)

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "Qwen/Qwen2.5-Coder-7B-Instruct"  # load compatibility only
REPO = "cristianoaredes/orqo-hype-v1"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, REPO, subfolder="lora")
model.eval()

2) Load merged (no PEFT)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

REPO = "cristianoaredes/orqo-hype-v1"
tok = AutoTokenizer.from_pretrained(REPO, subfolder="merged")
model = AutoModelForCausalLM.from_pretrained(
    REPO,
    subfolder="merged",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

3) GGUF (llama.cpp)

hf download cristianoaredes/orqo-hype-v1 \
  --include "gguf/*Q4_K_M.gguf" --local-dir ./orqo-hype-v1

./llama-cli -m ./orqo-hype-v1/gguf/sft-v9b-mix-Q4_K_M.gguf -p "5s logo sting..."

4) MLX (Apple Silicon)

hf download cristianoaredes/orqo-hype-v1 \
  --include "mlx-4bit/*" --local-dir ./orqo-hype-v1

python -m mlx_lm generate --model ./orqo-hype-v1/mlx-4bit \
  --prompt "5s logo sting navy gold fintech" --max-tokens 512

Verified export metadata: 4-bit quantization, group size 64, 4.501 bits/weight, converted on Apple Silicon.

Local benchmark recorded on 2026-07-17 with a local copy of the published MLX export, an Apple M5 Pro with 24 GB unified memory, mlx-lm 0.31.3, prompt length 512, 64 generated tokens and two measured trials after warmup:

Batch	Prompt tok/s (aggregate)	Generation tok/s (aggregate)	Peak memory (GB)
1	567.758	43.856	4.848
2	777.954	63.143	5.105
4	864.624	75.811	5.114
8	952.250	86.278	5.546

Method: the official synthetic python -m mlx_lm.benchmark runner. These are local throughput measurements for this exact configuration, not latency or capacity guarantees for another machine, context length or real creative workload. No speculative draft model was used.

5) Core ML

• Artifact: coreml/model.mlpackage.
• Recipe: static trace with seq_len=512, FP16 followed by 4-bit k-means palettization with group size 16.
• Deployment specification: iOS 18 / macOS 15 compatibility; runtime smoke was performed on macOS only.
• Limitation recorded by the exporter: no stateful KV cache, full forward pass per token, smoke-grade throughput, quality not benchmarked, and no physical-iOS-device validation.

---

Intended use

In scope

• Short motion-graphics / promo compositions for HyperFrames:
  • logo stings, title cards, product teasers
  • lower-thirds, reels-style kinetic type
  • multi-scene storyboards with GSAP timelines
• Staged outputs: design.json → storyboard.json → per-scene HTML

Out of scope

• Open-domain chat or general coding assistant
• Unfiltered web agent / tool-use
• Claiming pure single-shot quality equal to the full delivery pipeline (see Evaluation)

Recommended integration path

1. Staged generation (design → storyboard → scenes) — system prompts A/B/C
2. Optional deterministic post-decode assists (font inject, structure guards, motion-craft boost) when shipping quality floors
3. hyperframes assemble → lint
4. Rendering infrastructure is integrator-controlled and outside this model card's scope

Raw single-shot model quality ≠ full delivery-path metrics.

---

Prompting (staged)

The generation workflow uses three stages (compact form):

Stage	Role	Output
A	Art director	design.json — format, palette, typography, motion, style
B	Scene director	storyboard.json — scenes with duration + beat
C	Motion designer	Full HTML scene: GSAP CDN, data-composition-*, window.__timelines

Example brief:

8s product teaser: AI video editor, bold kinetic type, dark studio, navy and electric cyan

Versioned integration contract

The reference integration uses a single non-destructive tool named generate_hyperframe_composition. It accepts a bounded brief, target format, maximum scene count, temperature and output budget, then returns design, storyboard and structurally validated scene HTML under schema version 1.0.0; downstream HyperFrames lint is still required.

Input	Contract
brief	3–4000 characters
format	16:9, 9:16 or 1:1
max_scenes	integer from 1 to 8
temperature	number from 0 to 1.5
max_output_tokens	integer from 256 to 4096

This is an integration contract, not a hosted tool or endpoint. Backend URLs and credentials are operator configuration and are not accepted as tool arguments.

---

Training summary

Item	Value
Method	SFT adapter (sft-v9b-mix)
Train records	2150 (records_v9b_train, recipe-3 mix)
Epochs	2
Max sequence	8192
Learning rate	2e-4
Steps	538
Train loss	0.3317
Training hardware	NVIDIA A100-80GB

Subsequent experiments (motion DPO, atomic DPO, residual curriculum SFT) were not promoted over this champion.

Evidence: .archagents/13-execution/runs/RUN-20260711-des0012-e2e.md.

---

Evaluation (honest)

HyperFrames golden set n=72, Orqo delivery stack around this adapter (2026-07-11):

Gate	Result	Note
Lint (guarded delivery path)	72/72 (100%)	Deterministic post-decode guards included
Motion craft mean (uplifted delivery path)	7.0972	Temporal VLM evaluation on rendered MP4s
Diversity-out (student fonts)	soft PASS	Entropy 3.8629 after the font-pool assist
Raw single-shot lint	66/72 (91.7%)	Champion adapter without claiming the guarded pipeline as raw model quality

Evidence:

• .archagents/14-verify/reports/VER-served-sft-v9b-mix.json
• .archagents/14-verify/reports/VER-20260711-motion-craft-sft-v9b-mix-gate7.json
• .archagents/14-verify/reports/VER-20260711-diversity-out-sft-v9b-mix.md
• .archagents/14-verify/reports/VER-20260711-des0012-u61.md

---

Limitations

• Specialist model — weak outside HyperFrames composition briefs.
• Storyboard may emit source.type=block; those scenes are resolved from the deterministic allowlisted block catalogue and do not trigger a Stage C HTML request.
• Quality floors assume the delivery pipeline, not bare chat alone.
• MLX and GGUF quantisation may change creative output relative to merged weights.
• The Core ML export is smoke-grade: static sequence length, no stateful KV cache, no recorded quality benchmark, and no physical iOS device proof.
• No hosted endpoint, uptime commitment, or managed inference service is documented by this card.

---

License & attribution

• This packaging / adapter card: Apache-2.0
• Base model: follow Qwen/Qwen2.5-Coder-7B-Instruct terms
• HyperFrames / Orqo branding remain with their respective owners

---

Citation

@misc{orqo-hype-v1,
  title        = {Orqo Hype v1: HyperFrames Composition Specialist},
  author       = {Aredes, Cristiano},
  year         = {2026},
  howpublished = {Hugging Face model card},
  note         = {Public package: LoRA, merged weights, GGUF, MLX 4-bit, and Core ML},
  url          = {https://huggingface.co/cristianoaredes/orqo-hype-v1}
}

---

Changelog

Version	Date	Notes
orqo-hype-v1	2026-07-11	Branded packaging of champion sft-v9b-mix LoRA
orqo-hype-v1	2026-07-12	Initial package: lora/ · merged/ · gguf/
orqo-hype-v1	2026-07-17	Public inventory verified with MLX/Core ML metadata; hosted-serving references removed