Downloads · 30 days
19
4% of all-time downloads
Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16-v2
K-EXAONE-152B-A23B-REAP-80E-BF16-v2 is a text generation model from Baekpica. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as other.
This is a conservatively named v2 of Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16. The parent is a REAP-pruned version of LGAI-EXAONE/K-EXAONE-236B-A23B, with 80 routed experts per MoE layer and 23B active parameters.
Downloads · 30 days
19
4% of all-time downloads
All-time downloads
472
Public
Parameters
152B
304 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors304 GB · 100%
How the weights are stored.
BF16152B · 100%
From the Hugging Face model README
This is a conservatively named v2 of Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16. The parent is a REAP-pruned version of LGAI-EXAONE/K-EXAONE-236B-A23B, with 80 routed experts per MoE layer and 23B active parameters.
v2 applies a small post-pruning recovery run. It is not described as fully healed: measured gains are modest, and reasoning length remains sensitive to the generation cap.
The run used BF16 on four A100 80GB GPUs. Router load-balancing loss and router z-loss were included. No teacher model or distillation was used.
Small smoke test, not a general benchmark:
temperature=1.0, top_p=0.95, max_tokens=3072| Model | Closed </think> | Emitted EOT | Hit 3,072-token cap |
|---|---|---|---|
| Parent BF16 | 28/36 | 27/36 | 9/36 |
| v2 | 27/36 | 27/36 | 9/36 |
The held-out assistant CE improved from 1.18853 to 1.18012 and perplexity from 3.28226 to 3.25477. There were zero dead routed experts before and after.
Two recovery candidates were compared. A longer 110-step SFT candidate closed
</think> more often, but produced more tokens and higher repeated n-gram ratios.
This 50-step candidate was selected as the more conservative trade-off:
| Candidate | Generated tokens | Repeated 2-gram | Repeated 4-gram | Repeated 8-gram |
|---|---|---|---|---|
| 110-step | 77,484 | 0.5166 | 0.2405 | 0.0732 |
| v2 / 50-step | 74,862 | 0.5119 | 0.2343 | 0.0680 |
The full integrity and generation records are included as
integrity_report.json, staged_healing_report.json, and
evaluation_vllm024_3072.json.
transformers >= 5.1.0 is required for native exaone_moe support.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
dtype="bfloat16",
device_map="auto",
)
vllm serve Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16-v2 \
--tensor-parallel-size 4 \
--reasoning-parser deepseek_v3 \
--speculative-config '{"method":"mtp","num_speculative_tokens":2}'
Recommended starting sampling parameters are temperature=1.0 and top_p=0.95.
Do not interpret a 2K generation cap as proof of non-termination for this model;
the smoke tests showed materially more normal endings at a 3K cap.
이 모델은
Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16에
소규모 post-pruning recovery를 적용한 v2입니다. 완전한 복구를 의미하는
HEALED 명칭은 사용하지 않았습니다.
적용한 절차는 다음과 같습니다.
vLLM 0.24.0, 4×A100, 3개 seed, 36회 생성의 3,072-token 비교에서 부모와 v2는 EOT 27/36, cap 도달 9/36으로 동률이었습니다. Holdout CE는 1.18853에서 1.18012로 개선됐습니다. 더 긴 110-step 후보보다 반복 n-gram과 전체 생성량이 낮아 50-step 후보를 v2로 선택했습니다.
개선 폭은 제한적이며 모든 prompt에서 일관된 향상을 보장하지 않습니다. 특히
reasoning 길이는 max_tokens에 민감하므로 2K 제한만으로 무한 reasoning 여부를
판단하지 않는 것을 권장합니다. 실제 배포 전 한국어, reasoning, tool use 및
long-context 태스크를 별도로 검증하세요.
This model is distributed under the same K-EXAONE AI Model License as the
original model. See LICENSE for the complete terms.