pt-es-hi-sft

A 138M-parameter (205M including embeddings) decoder-only GPT for Portuguese, Spanish and Hindi, instruction-tuned from a pretrained checkpoint. Trained from scratch as a teaching/research exercise — it is small, and the limitations below are the interesting part of this card.

This is not a transformers model. The architecture (learned positional embeddings and qk-norm) matches no stock HF class, so AutoModel.from_pretrained will not load it. Load it against the training repo:

import json, torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from src.model.gpt import GPT, GPTConfig          # from the training repo
from transformers import AutoTokenizer

repo = "andre15silva/pt-es-hi-sft"
cfg = json.load(open(hf_hub_download(repo, "config.json")))
model = GPT(GPTConfig(**{k: cfg[k] for k in
    ("vocab_size", "block_size", "n_layer", "n_head", "n_embd", "dropout", "bias", "qk_norm")}))
model.load_state_dict(load_file(hf_hub_download(repo, "model.safetensors")), strict=False)  # lm_head is tied
model.eval()

tok = AutoTokenizer.from_pretrained(repo)
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Olá! Como estás?"}], tokenize=False, add_generation_prompt=True)
ids = torch.tensor([tok(prompt, add_special_tokens=False)["input_ids"]])
print(tok.decode(model.generate(ids, max_new_tokens=80, temperature=0.7, top_k=40)[0]))

strict=False is required: lm_head.weight is tied to transformer.wte.weight and is therefore not stored separately. GPT.__init__ re-ties them.

Prompt format

Always use the chat template. The model was fine-tuned on it exclusively and degrades on bare text, leaking <|eot_id|><|start_header_id|> into its output. Turns end with <|eot_id|>, not the tokenizer's <|end_of_text|> — stop generation on both.

Training

Loss is computed on assistant tokens only; every other position is masked to -1. One conversation per batch row, right-padded to the longest in the batch, with length-grouped batching (98.4% of the padded tensor was real tokens).

Data mix

By language: pt 44.9%, hi 35.3%, es 19.6%.

48.7% of rendered conversations (154,049 of 316,124) were dropped for exceeding the 1,024-token context window — a hard ceiling from the base model's learned positional embeddings. The filter is not uniform: magpie_ultra, the multi-turn general-instruction source, lost 88% of its rows; smol_rewrite lost 2.

Evaluation

Base checkpoint vs this model, n=500 per task, zero-shot. "chat" is the chat template; "raw" is plain ICL text.

Reading these honestly:

• PT-Culture (+51% over base in chat mode) is the real gain, and it beats the control: applying the chat template to the base model only reached 0.193, so most of the gain is the tuning, not the format.
• CALAME-PT went up (0.408 → 0.422), so fine-tuning did not damage raw language modelling — no catastrophic forgetting.
• The multiple-choice tasks are flat by design. They score by ranking fixed continuations by log-probability; SFT teaches format and stopping, not knowledge. Six of the eight tasks above are of this kind.
• EsCoLA is excluded: all four runs are degenerate (MCC ≈ 0 against accuracies swinging 0.32–0.69 on a ~2/3-imbalanced label, i.e. a near-constant predicted class).
• ALBA, the open-ended instruction-following benchmark, was not scored (no judge API key). It is the measurement most relevant to an SFT model and it is missing.

Limitations

It behaves like an assistant. It does not know very much.

"Olá! Como estás?"            → "Olá! Estou bem, obrigado por perguntar.
                                 E você, como está? 😊"      ✓ replies, stops
"Qual é a capital de França?" → "A capital de França é Paris."  ✓
"Qual é a capital de Portugal?" → "A capital de Portugal é Porto de Ave."  ✗

Lisboa never appeared across five seeds. The base model can complete "Lisboa é a capital de …" → "Portugal" correctly, but neither checkpoint answers the question form — a reversal-curse failure inherited from pretraining, not caused by SFT. Fine-tuning made confident fabrication more fluent: wrong answers arrive in the same assured, markdown-formatted register as correct ones.

Other known limitations:

• Hindi is weak. 35.3% of training tokens, but 94% of that is two templated generation tasks (headline→article, article→summary), not instruction following. Hindi chat replies are frequently incoherent where pt/es are clean, and ChatRAG-Hi is the one generative benchmark that got worse.
• No English. Neither pretraining nor SFT included it. English prompts are usually answered in Spanish or Portuguese.
• 1,024-token context, hard-capped by learned positional embeddings.
• No safety tuning of any kind. No RLHF, no DPO, no refusal training, no red-teaming. Do not deploy this.
• Untuned hyperparameters. lr 5e-5 / 5e-6 is the conventional ~10×-below-pretraining starting point, not a swept optimum. Validation loss was still falling at 2 epochs, so the model is undertrained.

Reproducing

uv run scripts/prepare_sft_data.py --tokenizer-dir <tokenizer> \
    --out-dir data/sft --max-example-tokens 1024
uv run scripts/train_sft.py --init-from <pretrain ckpt> --data-dir data/sft \
    --out-dir runs/sft --block-size 1024 --batch-size 16 --max-iters 19854 \
    --lr 5e-5 --min-lr 5e-6 --wandb

See scripts/slurm/13_prepare_sft_data.sh and 14_train_sft.sh in the training repo for the cluster versions with the derived iteration counts.