Key Specifications
| Vendor | other |
|---|
| Version | falcon-40b |
|---|
| Release Date | 2023-05-25 |
|---|
| Context Window | 2048 tokens |
|---|
| Input Modalities | text |
|---|
| Output Modalities | text |
|---|
| License | TII Falcon LLM License |
|---|
| Documentation | https://huggingface.co/models |
|---|
Benchmark Performance
| Benchmark | Score | Unit | Evaluated At | Notes | Source |
|---|
| MMLU | 56.7 | % | 2023-05-25 | 5-shot | view |
| HUMANEVAL | 39.7 | pass@1 | 2023-05-25 | — | view |
| GSM8K | 61.3 | % | 2023-05-25 | 0-shot CoT | view |
| MATH | 21.3 | % | 2023-05-25 | 0-shot CoT | view |
| BBH | 50.6 | % | 2023-05-25 | 3-shot CoT | view |
| GPQA | 21.1 | % | 2023-05-25 | 0-shot | view |
| IFEVAL | 57.2 | % | 2023-05-25 | prompt_strict | view |
| ARC | 82.9 | % | 2023-05-25 | challenge | view |
| MUSR | 38.1 | % | 2023-05-25 | 0-shot | view |
| WINOGRANDE | 65.6 | % | 2023-05-25 | 0-shot | view |
Pricing
| Tier | Price | Currency |
|---|
| Input | $0.4 / Mtok | USD |
| Output | $0.4 / Mtok | USD |
| Cache Read | $0 / Mtok | USD |
| Cache Write | $0 / Mtok | USD |
Source:
https://huggingface.co/models
· as of 2023-05-25
Compliance
- Data Residency: self-host
- SOC2: ✗
- HIPAA: ✗
- GDPR: ✗
- ISO 27001: ✗
Falcon 40B
Ikhtisar Model
TII Falcon 40B 中型开源模型, 2K 上下文, 400 亿参数, 在 40B 规模上曾领先 Llama 2 70B。
Spesifikasi Inti
| Vendor | Versi | Tanggal Rilis | Jendela Konteks | Modalitas Input | Modalitas Output | Lisensi |
|---|
| Other | falcon-40b | 2023-05-25 | 2K | text | text | TII Falcon LLM License |
Kinerja Benchmark
| Benchmark | Skor | Satuan | Catatan |
|---|
| MMLU (Massive Multitask Language Understanding) | 56.7 | % | 5-shot |
| HumanEval | 39.7 | pass@1 | — |
| GSM8K (Grade School Math 8K) | 61.3 | % | 0-shot CoT |
| MATH | 21.3 | % | 0-shot CoT |
| BBH (BIG-Bench Hard) | 50.6 | % | 3-shot CoT |
| GPQA | 21.1 | % | 0-shot |
| IFEval | 57.2 | % | prompt_strict |
| ARC | 82.9 | % | challenge |
| MUSR | 38.1 | % | 0-shot |
| WinoGrande | 65.6 | % | 0-shot |
Harga
| Input | Output | Baca Cache | Tulis Cache |
|---|
| — | — | — | — |
per juta token
Kelebihan
Kekurangan
- MMLU 仅 56.7，知识推理偏弱。
- HumanEval 39.7，代码能力较弱。
- 闭源专有模型，不支持自托管。
- 上下文窗口 2K 偏小。
Kasus Penggunaan
Referensi