Qwen3.8-4B-A1-DARE-TIES
Configuracao do Merge
| Parametro | Valor |
|---|---|
| Metodo | DARE-TIES |
| Base | Qwen/Qwen3.5-4B |
| Fonte 1 | empero-ai/Qwen3.8-4B-Distill (weight=0.7) |
| Fonte 2 | InternScience/Agents-A1-4B (weight=0.3) |
| Density | 0.7 |
| Seed | 42 |
| Shards | 2 (5.01 GB + 3.67 GB) |
| Total | 8.68 GB (bf16) |
Objetivo do Merge
Combinar a deducao logica do Qwen3.8-4B-Distill (78% em BBH Logical Deduction) com a capacidade agentica do Agents-A1-4B (38% em BBH Causal Judgement).
Benchmarks dos Modelos Pais
| Modelo | Deducao | Causal | Medicina |
|---|---|---|---|
Qwen3.8-4B-Distill | 78% | 46% | 88% |
Agents-A1-4B | 26% | 38% | 88% |
| Este merge (0.7/0.3) | a medir | a medir | a medir |
Uso
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
)
model = AutoModelForCausalLM.from_pretrained(
"leandronunes/qwen3.8-4b-a1-dare-ties",
quantization_config=bnb_config,
device_map="auto",
trust_remote_code=True,
attn_implementation="sdpa",
)
tokenizer = AutoTokenizer.from_pretrained(
"leandronunes/qwen3.8-4b-a1-dare-ties",
trust_remote_code=True,
)
Notas Tecnicas
- Modelo multimodal (VLM): inclui torre de visao de 24 blocos alem do language model.
- Arquitetura hibrida: 3 camadas
linear_attention(Gated DeltaNet) seguidas de 1full_attention, repetido 8 vezes. - Thinking mode: para desativar o raciocinio estendido, use
enable_thinking=Falsenoapply_chat_template. - Contexto: 262.144 tokens (256K).
Licenca
Apache 2.0 (herdada do Qwen3.5-4B).