Hatsunama/Qwen3-4B-OBLITERATED-toolcall-lora

Qwen3-4B-OBLITERATED tool-calling LoRA

Short SFT LoRA so the repaired OBLITERATED 4B emits Qwen3 native tool calls:

<tool_call>
{"name": "fn", "arguments": {...}}
</tool_call>

Load this on [`Hatsunama/Qwen3-4B-OBLITERATED-repaired`](https://huggingface.co/Hatsunama/Qwen3-4B-OBLITERATED-repaired), not on `OBLITERATUS/Qwen3-4B-OBLITERATED` (NaN embeddings) and not on the Q4_K_S GGUF. Stop generation on </tool_call>.

Results

Greedy decode, Qwen3 tools= in the chat template.

Leftover miss: unbound delete_database is still emitted. BFCL numbers use a simplified AST check, not the official scorer.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Hatsunama/Qwen3-4B-OBLITERATED-repaired"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "Hatsunama/Qwen3-4B-OBLITERATED-toolcall-lora")

llama.cpp file: Qwen3-4B-OBLITERATED-toolcall.Q4_K_S.gguf on `Hatsunama/Qwen3-4B-OBLITERATED-Q4_K_S`.