metafloor-ai/manifest-specialist-demand-planning-v0.6.0

Manifest Specialist — Demand Planning & Forecasting · v0.6.0

A specialized supply-chain domain-expert — focused entirely on demand planning & forecasting.

Manifest Demand Planning is a domain-expert in the Manifest family: where the general Manifest orchestrator models cover all of supply chain, this one is tuned specifically for demand planning and forecasting — baseline vs promotional uplift, forecast methods, seasonality, and forecast accuracy.

• —Family: Manifest · Type: domain-expert (specialized) · Size: 2B
• —Specialty: demand planning & forecasting

### Preferred 82.5% of the time over the base model — 85.0% with matched format On a focused 20-question demand planning benchmark, an independent LLM judge panel preferred this model's answer over the base model's 82.5% of the time (16 / 3 / 1 W/L/T), and 85.0% with the same answer format. (20 items is a small sample — treat exact figures as directional.)

What's new in v0.6

• —Retrained on the expanded v0.6 supply-chain data — this specialist's focused split now draws from MetaFloor's ~32k-example corpus (~4k in-domain examples).
• —Scored on the current 20-question demand planning benchmark; the 35B orchestrator is retired this release and a new 9B tier is coming.
• —This is a new versioned repo — the previous v0.5 release (`metafloor-ai/manifest-specialist-demand-planning`, unchanged) keeps its own history and download stats.

v0.5 → v0.6 (preferred over the untuned base, same benchmark):

Best for

Demand planning and forecasting questions where a focused, practitioner's answer beats a generic one:

• —Baseline vs promotional-uplift separation and event forecasting
• —Choosing and tuning forecast methods (moving average, exponential smoothing, etc.)
• —Seasonality, forecast accuracy (MAPE/bias) and demand-sensing trade-offs

See the difference

Same question. Base model vs this specialist.

Ask (supply-chain director, mostly manual planning, prepping a trade promotion): "Why separate baseline demand from promotional uplift?" Base model → "…the separation of baseline demand from promotional uplift is critical for this mid-market scenario due to the company's reliance on a simple moving average…" — hedged and generic. Manifest Demand Planning → "A moving average of total history lumps every past period together, so it smears a one-time promo into a baseline and then over-forecasts the weeks after the event…" — direct, mechanism-first, practitioner framing.

The Manifest family

Two kinds of models:

🧭 Orchestrators — general-purpose, handle any supply-chain area

🎯 Domain-experts — specialized for a single area

Orchestrators are scored on the general supply-chain benchmark; domain-experts on their focused domain benchmark (20 items each — treat those exact figures as directional).

How to use

This is a LoRA adapter (~44 MB), applied on top of its base model at load time.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3.5-2B"  # base model — see "Built on" below
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "metafloor-ai/manifest-specialist-demand-planning-v0.6.0")

SYSTEM = "You are a senior supply chain expert. Answer correctly and concisely."
msgs = [{"role": "system", "content": SYSTEM},
        {"role": "user", "content": "How should I forecast a new SKU with only 8 weeks of history and a promo coming up?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_dict=True, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Prompt tip: include the asker's role and operating scale (revenue, SKUs, suppliers, nodes, lead time) for the sharpest answers.

Training details

The training data is a focused, proprietary demand-planning dataset and is not open-sourced — only the held-out evaluation benchmark (supply-chain-eval) is public.

Evaluation

Scored on a focused 20-question demand-planning benchmark — pairwise LLM-as-judge (2-model panel), this model's answer vs the base model's for the same prompt.

Benchmark: [supply-chain-eval](https://huggingface.co/datasets/metafloor-ai/supply-chain-eval) (planning split). The sample is small (20 items), so treat exact figures as directional.

Intended use & limitations

• —Intended use: decision-support and drafting for demand planning & forecasting questions.
• —Out of scope: not legally binding, contractual, or safety-critical guidance; no access to your live systems. Verify outputs before acting.
• —Limitations: English-only; specialized to demand planning (use a Manifest orchestrator for general questions); trained on synthetic data; standard LLM risks apply.

License

Released under CC-BY-NC-4.0 — free for research and non-commercial use, with attribution. Commercial use requires a license from MetaFloor — get in touch at metafloor.ai.

Built on

This model is a LoRA adapter over Qwen/Qwen3.5-2B (used under its own license); the base model is required to load the adapter.

Citation

@misc{metafloor_manifest_demand_planning,
  title  = {Manifest Specialist — Demand Planning: a demand-planning domain-expert (MetaFloor Manifest family)},
  author = {MetaFloor AI},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/metafloor-ai/manifest-specialist-demand-planning-v0.6.0}}
}