---
license: apache-2.0
base_model: Ilides/coser-1.1-code
tags:
  - coser-1.3-coder
  - ilides
  - coding-assistant
  - agentic
  - qwen3.5
  - lora
language:
  - en
  - es
pipeline_tag: text-generation
---

# Coser 1.3-coder by ilides (HF Safetensors)

**Coser 1.3-coder** es el asistente de código agéntico de **ilides**: identidad Coser clara, tono natural (con humor ligero), y enfoque en ingeniería real — planificar, depurar y escribir código de producción.

Evolución de [Coser 1.1-code](https://huggingface.co/Ilides/coser-1.1-code), fine-tuned con **74 ejemplos curados** (código, identidad, chat y correcciones de comportamiento).

Publicado por **[ilides](https://huggingface.co/Ilides)**.

## Versiones

| Repositorio | Formato | Uso |
|-------------|---------|-----|
| [Ilides/coser-1.3-coder](https://huggingface.co/Ilides/coser-1.3-coder) | Safetensors | Transformers, fine-tuning |
| [Ilides/coser-1.3-coder-GGUF](https://huggingface.co/Ilides/coser-1.3-coder-GGUF) | GGUF F16 + Q8_0 | llama.cpp, LM Studio |

## Identidad

System prompt recomendado:

```
You are Coser 1.3-coder by ilides, an expert AI coding assistant. Always speak in first person. Never say the user is Coser. Use natural prose unless the user explicitly asks for JSON.
```

## Stats de entrenamiento

| Métrica | Valor |
|---------|-------|
| Base | Coser 1.1-code (Qwen3.5-0.8B) |
| Dataset | 74 ejemplos curados |
| Método | LoRA r=16 + QLoRA 4-bit |
| Steps | 95 |
| Épocas | 5 |
| Loss final | 0.5570954799652099 |
| Token accuracy | 89.5% |
| Tiempo | 13.7 min |
| GPU | NVIDIA GeForce RTX 3050 |

## Benchmark (NVIDIA GeForce RTX 3050)

| Prompt | tok/s |
|--------|-------|
| Write a Python function that reverses a linked l... | 18.9 |
| Write a JavaScript async function to fetch and p... | 19.9 |
| Explain what binary search is and write it in Py... | 16.2 |
| Write a SQL query to find duplicate emails in a ... | 16.3 |
| Fix this bug: my Python function returns None in... | 16.2 |
| **Promedio** | **17.5** |

## Uso (Transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Ilides/coser-1.3-coder"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", trust_remote_code=True, torch_dtype=torch.bfloat16
)
messages = [
    {"role": "system", "content": "You are Coser 1.3-coder by ilides, an expert AI coding assistant."},
    {"role": "user", "content": "Who are you?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, temperature=0.55, repetition_penalty=1.12)
print(tokenizer.decode(out[0], skip_special_tokens=True))
```

## GGUF (llama.cpp)

```bash
llama-cli -m coser-1.3-coder-q8_0.gguf -cnv -ngl 99
```

## Créditos

- Base: [Ilides/coser-1.1-code](https://huggingface.co/Ilides/coser-1.1-code)
- Autor: ilides