README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language:
- en
- ko
tags:
- darwin
- expert
- multi-domain
- task-arithmetic
- cohere2
- moe
base_model: FINAL-Bench/Darwin-218B-kr
---

# 🌌 Darwin-218B-Expert

> **VIDRAFT FINAL-Bench**
> Darwin family multi-domain Expert flagship — 218B parameter Cohere2 MoE base with chemistry + biology domain knowledge integrated via **task arithmetic** on Korean-aligned base.

A multi-domain "Expert" derivative of the Darwin-218B family. Built by composing chemistry and biology domain deltas onto the Korean-aligned base, this v1 release preserves Korean fluency while adding scientific reasoning in chemistry and biology.

---

## Lineage

```
CohereLabs/command-a-plus-05-2026-bf16     (base 218B MoE, ~25B active)
              ↓ Korean LoRA merge
       Darwin-218B-kr                       (Korean-aligned base)
              ↓ Chem LoRA merge        ↓ Bio LoRA merge
   Darwin-218B-Expert-Chem      Darwin-218B-Expert-Bio
                       ↓ task arithmetic ↓
              Darwin-218B-Expert  ← THIS MODEL (v1 = Chem + Bio)
```

**Construction (v1)**: task arithmetic on the Korean base
```
W_Expert = W_Chem + W_Bio − W_kr
        = W_kr + ΔChem + ΔBio
```

Both chemistry and biology deltas preserved at 100% (no dilution), with Korean fluency inherited from the kr base.

---

## Domain Capabilities

| Domain | Coverage | Source |
|--------|----------|--------|
| **Korean (한국어)** | Native fluency | Inherited from `Darwin-218B-kr` base |
| **Chemistry** | Organic, spectroscopy, physical, inorganic, analytical, special (6-domain SFT) | `Darwin-218B-Expert-Chem` delta (Opus-distilled, anti-contamination) |
| **Biology** | Cellular, molecular, biochem, ecology, evolution (6-domain SFT) | `Darwin-218B-Expert-Bio` delta (Opus-distilled, anti-contamination) |
| **General** | All capabilities from Cohere Command A+ retained | Underlying base |

---

## Architecture

| Item | Value |
|------|-------|
| Parameters | 218B total / ~25B active (MoE) |
| Architecture | `Cohere2VisionForConditionalGeneration` (multimodal-capable, text-primary) |
| Experts | 128 (Cohere2 MoE structure) |
| Precision | BF16 |
| Tokenizer | Cohere2 (vocab 256K) |
| Languages | English, Korean |
| Context | 65,536 tokens |

---

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "FINAL-Bench/Darwin-218B-Expert",
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-218B-Expert")

messages = [
    {"role": "user", "content": "Explain the mechanism of SN2 reaction step by step. Then describe how mRNA splicing maintains reading frame fidelity."}
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

**Serving (vLLM, recommended)**:
```bash
vllm serve FINAL-Bench/Darwin-218B-Expert \
    --tensor-parallel-size 8 \
    --dtype bfloat16 \
    --max-model-len 65536 \
    --trust-remote-code
```

Requires vLLM ≥ 0.21.0 with `Cohere2VisionForConditionalGeneration` support.

---

## Methodology — Task Arithmetic

Unlike sibling-model averaging (which dilutes expertise), this model uses **task vectors** (Ilharco et al., 2022) to compose domain knowledge:

1. Train domain-specific LoRA on a shared base (`Darwin-218B-kr`)
2. Each merged expert (`Expert-Chem`, `Expert-Bio`) encodes a domain delta from the base
3. Add deltas to the base: `W_final = W_base + Σ Δ_domains`

This is mathematically equivalent to applying multiple LoRA adapters simultaneously, but produces a single dense checkpoint requiring no runtime adapter loading.

**Properties**:
- Both expert deltas preserved at 100% magnitude
- No interference dilution (unlike weighted blends)
- Korean base layer intact

---

## Version Roadmap

| Version | Domains | Status |
|---------|---------|--------|
| **v1** | Korean + Chemistry + Biology | ✅ Current |
| v2 | + Mathematics | ⬜ Planned |
| v3 | + Code | ⬜ Planned |
| v4 | + Physics, Law | ⬜ Planned |

Each domain delta is added via task arithmetic on the shared `Darwin-218B-kr` base.

---

## License

**Apache License 2.0**

Built upon `CohereLabs/command-a-plus-05-2026-bf16` (Apache-2.0) and `Darwin-218B-kr` (Apache-2.0). Both upstream components are permissively licensed; users are responsible for compliance with upstream terms.

---

## Contributors

**Lead Architect & Developer**
**장재원 (Jaewon Jang)** — CTO, VIDRAFT
*Multi-domain Expert composition design, task-arithmetic merge pipeline, domain SFT distillation pipelines.*

**Organization**
VIDRAFT / FINAL-Bench
https://huggingface.co/FINAL-Bench

---

## Citation

```bibtex
@misc{darwin-218b-expert-2026,
  title  = {Darwin-218B-Expert: Multi-Domain Expert via Task Arithmetic on Korean-Aligned 218B MoE},
  author = {Jang, Jaewon and {VIDRAFT FINAL-Bench Team}},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-218B-Expert}}
}
```

---

## References

- **Task Arithmetic**: Ilharco et al., "Editing Models with Task Arithmetic", ICLR 2023 — [arXiv:2212.04089](https://arxiv.org/abs/2212.04089)
- **TIES-Merging**: Yadav et al., "TIES-Merging: Resolving Interference When Merging Models", NeurIPS 2023 — [arXiv:2306.01708](https://arxiv.org/abs/2306.01708)
- **Cohere Command A+** (base): [CohereLabs/command-a-plus-05-2026-bf16](https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16)
- **Darwin-218B-kr** (Korean base): private — `FINAL-Bench/Darwin-218B-kr`
- **Expert sources**: private — `FINAL-Bench/Darwin-218B-Expert-Chem`, `FINAL-Bench/Darwin-218B-Expert-Bio`