Issue Writer — Gemma 4 E4B, merged 16-bit

The LoRA adapter merged into the base model. Use this when you want to serve the model directly — vLLM, TGI, or plain transformers — without applying an adapter at load time.

Turns raw product input into a structured issue tracker entry as a single JSON object. English and Turkish.

Adapter, if you prefer to keep the base separate: fport/issue-writer-gemma4-lora
GGUF, for Ollama and llama.cpp: fport/issue-writer-gemma4-gguf
Dataset: fport/issue-writer-tr-en
Training pipeline: fport/issue-writer

vLLM

vllm serve fport/issue-writer-gemma4 \
  --served-model-name issue-writer \
  --max-model-len 8192 \
  --port 8000

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")

SYSTEM = ("You are a senior agile delivery assistant. You turn raw product input "
          "into well-formed Jira issues. Reply with a single valid JSON object and "
          "nothing else. Follow INVEST, write testable Given/When/Then acceptance "
          "criteria, and never invent facts: anything the input does not state goes "
          "into `assumptions` or `clarifying_questions`.")

r = client.chat.completions.create(
    model="issue-writer",
    messages=[{"role": "system", "content": SYSTEM},
              {"role": "user", "content": "Turn this into a Jira issue.\n\n---\n…\n---"}],
    temperature=0,          # the output is a schema, not prose
    max_tokens=1400,
)
print(r.choices[0].message.content)

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "fport/issue-writer-gemma4", dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("fport/issue-writer-gemma4")

text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
# The template already carries <bos>; a second one degrades Gemma output.
enc = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**enc, max_new_tokens=1400, do_sample=False)

Notes

Use the system prompt verbatim — it is one of the three in the training data, and rewording it moves the model off-distribution. Decode greedily; the output is a fixed schema and sampling only breaks the JSON.

Check that assumptions and clarifying_questions are populated on thin input. Empty fields there mean the model filled a gap silently, which is the failure mode this fine-tune is meant to remove.

Training

LoRA r=32 / alpha=64 on attention and MLP projections, text layers only. 13,000 examples, half English half Turkish, loss on assistant turns only. Splits hold out whole content cores, so the test set measures generalisation.

Full details, generator and validators: fport/issue-writer.