You can deploy this model and start earning money today!

jpalmae/
qwen35b-espanol-abl-merged-v3

Name: qwen35b-espanol-abl-merged-v3
Availability: InStock

safetensors

qwen3_5_moe

mit

Views

No views yet

qwen35b-espanol-abl-merged-v3

Modelo fusión (merge) multimodal Qwen3.5/3.6 35B-A3B entrenado para español neutro/chileno con filtrado anti-rioplatense.

Pipeline

Base: qwen35b-espanol-abl-merged-v2 (Qwen3.5 35B-A3B abliterated multimodal).
Dataset v3: 200k ejemplos (60k chilenos + 60k neutros + 80k filtrados anti-rioplatense).
Entrenamiento: LoRA r=64, alpha=128, 2.500 pasos, batch 1, accum 32, seq 256, bf16, sdpa, gradient checkpointing.
Loss final: ~2.11.
Merge: adaptador LoRA aplicado sobre la base multimodal completa (Qwen3_5MoeForConditionalGeneration).
Despliegue: vLLM en Puget (2x RTX PRO 6000 Blackwell).

Arquitectura

model_type: qwen3_5_moe
architectures: ["Qwen3_5MoeForConditionalGeneration"]
Multimodal: sí, mantiene vision_config y tokens de imagen/video.
Bf16, ~66 GB en disco.

Uso

python

from transformers import AutoModelForConditionalGeneration, AutoTokenizer
model = AutoModelForConditionalGeneration.from_pretrained(
    "jpalmae/qwen35b-espanol-abl-merged-v3",
    torch_dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "jpalmae/qwen35b-espanol-abl-merged-v3",
    trust_remote_code=True,
)