qwen3-4b-subtitle-es-v3-lora

qwen3-4b-subtitle-es-v3-lora

QLoRA adapter for Qwen3-4B that performs constrained EN→ES subtitle translation: given a scene of English subtitle lines (12–28 lines, one per line), it outputs exactly one Spanish line per source line — no lines added, dropped, or merged — with no output line exceeding 42 characters, in a single direct pass (no extended thinking).

How to run (measured configuration)

Important: load the adapter against Qwen/Qwen3-4B as shown below. The base_model_name_or_path in adapter_config.json points at a 4-bit Unsloth variant (a training-time artifact); loading against that base is not the configuration this model was evaluated in and will fail on machines without bitsandbytes.

The evaluated regime is: fp16, greedy decoding, thinking disabled, and the system prompt below verbatim.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

ADAPTER = "Hookem22/qwen3-4b-subtitle-es-v3-lora"

tok = AutoTokenizer.from_pretrained(ADAPTER)
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B", dtype=torch.float16)
model = PeftModel.from_pretrained(base, ADAPTER).eval()
device = ("cuda" if torch.cuda.is_available()
          else "mps" if torch.backends.mps.is_available() else "cpu")
model.to(device)

SYSTEM = ("Translate these subtitles from English to Spanish. Output exactly one "
          "Spanish line per source line, and keep every line to 42 characters or fewer.")

english_scene = """\
I need the report by Friday.
That's not enough time.
Then ask Marcus for help."""  # normally 12-28 lines

messages = [{"role": "system", "content": SYSTEM},
            {"role": "user", "content": english_scene}]
prompt = tok.apply_chat_template(messages, tokenize=False,
                                 add_generation_prompt=True,
                                 enable_thinking=False)
inputs = tok(prompt, return_tensors="pt").to(device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=1024, do_sample=False,
                         pad_token_id=tok.eos_token_id)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:],
                 skip_special_tokens=True).strip())

Output contract

Any output can be checked mechanically:

• line count of the output equals line count of the input
• every output line is ≤ 42 characters (wc -L ≤ 42)
• lines correspond 1:1 and in order to the source lines

Notes

• Input format: plain English subtitle lines separated by newlines in the user message — no numbering, no timestamps.
• The 42-character limit counts every character including spaces and punctuation.
• Trained with completion-only loss on scene-scale examples; the system prompt above is byte-identical to the one used in training.