Downloads · 30 days
9
10% of all-time downloads
axiom-of-choice/qwen3-1.7b-es-reasoning-peft
qwen3-1.7b-es-reasoning-peft is a text generation model from axiom-of-choice. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
A transformers/peft-compatible port of axiom-of-choice/qwen3-1.7b-es-reasoning-lora, a rank-8 LoRA adapter trained in MLX that makes Qwen/Qwen3-1.7B reason in Spanish. Same
Downloads · 30 days
9
10% of all-time downloads
All-time downloads
87
Public
Repo size
20 MB
Likes
1
Public
Click a slice to open those files.
.safetensors20 MB · 100%
From the Hugging Face model README
A transformers/peft-compatible port of
axiom-of-choice/qwen3-1.7b-es-reasoning-lora, a rank-8 LoRA
adapter trained in MLX that makes
Qwen/Qwen3-1.7B reason in Spanish. Same
adapter weights, converted losslessly into the shapes and names peft
expects -- every LoRA update matrix (B @ A) verified equal to the MLX
original at rtol=1e-5. If you use mlx-lm, get the original repo instead;
this one is for everyone else (transformers, vLLM with --enable-lora,
anything built on peft).
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", dtype="bfloat16")
model = PeftModel.from_pretrained(base, "axiom-of-choice/qwen3-1.7b-es-reasoning-peft")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")
messages = [{"role": "user", "content": "¿Cuánto es 17 por 24?"}]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
)
out = model.generate(inputs, max_new_tokens=1024)
print(tokenizer.decode(out[0], skip_special_tokens=True))
Format conversion (key names, tensor orientation, scale→lora_alpha) is
checked with no model load and proves the files are self-consistent, but not
that this repo behaves like the original. This table is a real generation
run: 100 items from the held-out Spanish GSM8K eval, this repo
(Qwen/Qwen3-1.7B, bf16) paired item-by-item against the original
(Qwen/Qwen3-1.7B, MLX -- same bf16 checkpoint, run with mlx-lm).
| accuracy | agreement | |
|---|---|---|
| this repo | 71.0% | 84.0% |
| original (Qwen/Qwen3-1.7B, MLX) | 79.0% |
Paired counts: 4 correct only here, 12 correct only in the original, out of 100. A measured, directional gap: 12 items only the original gets right against 4 only this port gets right. Some of that is decode-path noise (MLX and PyTorch/MPS use different numerical kernels; greedy decoding has zero tolerance for the resulting rounding differences, so a flipped argmax on one low-margin token can fork the rest of a generation) -- but not all of it measures as noise. Treat the numbers below as the actual behavior of this repo, not as equal to the original.
Only the adapter tensors are provably identical to the original. The base model is not always the same weights: MLX may run a quantized base (dequantized on the fly) where this repo runs bf16 natively, and MLX and PyTorch are different numerical engines even on identical weights. The table above is the actual measurement, not an assumption that identical adapter math implies identical behavior.
Same recipe, dataset and MLX-side evaluation as the original --
see axiom-of-choice/qwen3-1.7b-es-reasoning-lora for the full
training card.
Puerto a transformers/peft de
axiom-of-choice/qwen3-1.7b-es-reasoning-lora, un adaptador LoRA de
rango 8 entrenado en MLX que hace que
Qwen/Qwen3-1.7B razone en español.
Mismos pesos del adaptador, convertidos sin pérdida al formato que espera
peft -- cada matriz de actualización LoRA verificada igual al original MLX
con rtol=1e-5.
| precisión | acuerdo | |
|---|---|---|
| este repo | 71.0% | 84.0% |
| original (Qwen/Qwen3-1.7B, MLX) | 79.0% |
Conteos pareados: 4 correctos solo aquí, 12 correctos solo en el original, de 100. Una brecha real y direccional: 12 ítems donde solo el original acierta contra 4 donde solo acierta este port. Parte es ruido del camino de decodificación (MLX y PyTorch/MPS usan kernels numéricos distintos; la decodificación greedy no tolera esas diferencias de redondeo, así que un argmax volteado en un token de bajo margen puede desviar el resto de la generación) -- pero no todo mide como ruido. Trata los números de abajo como el comportamiento real de este repo, no como equivalentes al original.
Misma receta de entrenamiento y evaluación que el original -- ver
axiom-of-choice/qwen3-1.7b-es-reasoning-lora para la ficha
completa.