---
license: llama3.2
base_model: meta-llama/Llama-3.2-1B
language:
- es
- pt
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- sft
- latam
- llama
- latam-gpt
datasets:
- latam-gpt/wikipedia-qa-iberoamerica-clean
- latam-gpt/tulu-3-sft-mixture-no-identity-clean
- latam-gpt/sft-mcqa-unified-clean
- latam-gpt/self-identity-v3-clean
- latam-gpt/copuchatV2-clean
---

# latamgpt-1b-sft

Full supervised fine-tune of `meta-llama/Llama-3.2-1B` on a Latin American instruction mixture, with sequence packing and assistant-only loss masking.

## Training data

| Dataset | Role |
|---|---|
| `latam-gpt/wikipedia-qa-iberoamerica-clean` | Iberoamerican QA |
| `latam-gpt/tulu-3-sft-mixture-no-identity-clean` | General SFT mixture |
| `latam-gpt/sft-mcqa-unified-clean` | Multiple-choice QA |
| `latam-gpt/self-identity-v3-clean` | Self-identity |
| `latam-gpt/copuchatV2-clean` | Chilean conversational |

After tokenization with the Llama 3 chat template and FFD packing at 4096 tokens (99.16% packing efficiency), the training set is **175,995 packed sequences (~721M tokens)** with **119,573 validation examples**.

## Training setup

- **Base model:** `meta-llama/Llama-3.2-1B`
- **Method:** Full fine-tune (no LoRA)
- **Sequence length:** 4096, packed with assistant-only loss masking
- **Epochs:** 2
- **Learning rate:** 1e-5 (cosine, 3% warmup)
- **Batch:** 8 per device × 8 grad accumulation × 2 GPUs = **128 effective**
- **Precision:** bf16 + DeepSpeed ZeRO-2 + gradient checkpointing + Flash Attention 2
- **Hardware:** 2× H100 80GB
- **Wall time:** ~7.5h

## Results

| Metric | Value |
|---|---|
| `train_loss` | 0.8579 |
| `eval_loss` | 0.8910 |
| `epoch` | 2.0 |
| `train_samples` | 175,995 |

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("latam-gpt/latamgpt-1b-sft")
model = AutoModelForCausalLM.from_pretrained("latam-gpt/latamgpt-1b-sft", torch_dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "¿Cuál es la capital de Chile?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
```

## License

Inherits the [Llama 3.2 Community License](https://www.llama.com/llama3_2/license/) from the base model.