sixfingerdev/sixfinger-phi2-merged

SixFinger Phi-2 — standalone experimental checkpoint

SixFinger Phi-2 is a standalone ~2.7B-parameter causal language model checkpoint with a Microsoft Phi-2 architecture. It is published as full model weights, rather than a separate LoRA adapter.

Developer: Six Finger Dev (Enes Altıparmak)

Provenance and evaluation

The project was described as a Turkish fine-tuning/adapter-merging experiment based on Microsoft Phi-2. However, the current repository does not include sufficient training logs, a dataset manifest, or adapter history to independently reproduce that process. We therefore do not claim a measured improvement over Phi-2.

The included model report describes local observations:

• The checkpoint uses a standard PhiForCausalLM architecture with full weights.
• It has no defined chat template and behaves more like a completion model than a chat assistant.
• Results on a few local prompts were mixed; they are not a representative benchmark.
• Turkish conversation, repetition control, and instruction following may be unreliable.

Installation

pip install -U torch transformers accelerate

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sixfingerdev/sixfinger-phi2-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
model.eval()

prompt = "Soru: Türkiye'nin başkenti neresidir?\nCevap:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=80,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = outputs[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Limitations

• This is not a validated production assistant.
• No reproducible evaluation against the base model is currently published.
• A prompt-completion format with explicit cues (such as Cevap:) may work better than chat-style messages.
• Outputs may contain factual errors or repetition. Verify results for consequential uses.

License

This checkpoint is published under MIT, consistent with the Microsoft Phi-2 model license. Third-party training data, if used, may have separate terms.