Clarify model limitations and fix usage instructions

README.md

@@ -7,7 +7,7 @@ tags:
 - lora
 - sft
 - transformers
-- trl
+- peft
 - unsloth
 license: apache-2.0
 datasets:
@@ -15,108 +15,73 @@ datasets:
 language:
 - tr
 - en
-- zh
 ---
-# SixFinger-8B Adapter for LLaMA 3.1 8B  
 
-This repository contains a **LoRA adapter** for the SixFinger-8B model.  
-The adapter allows fine-tuned responses on top of the base model **```unsloth/llama-3.1-8b-bnb-4bit```** without modifying the base weights.  
+# SixFinger-8B — Llama 3.1 8B LoRA adapter
 
----
+**This repository contains a LoRA adapter, not a standalone 8B model.** To run it, you need the original Llama 3.1 8B base model and the PEFT library. This is an experimental Turkish-language fine-tuning project.
 
-## Overview
+## What is included
 
-- **Base Model:** unsloth/llama-3.1-8b-bnb-4bit  
-- **Adapter Type:** LoRA  
-- **Quantization:** 4-bit (via bitsandbytes)  
-- **Purpose:** Enhanced response generation for Turkish/English mixed datasets.  
-- **Compatibility:** Use with Hugging Face Transformers + PEFT library.  
+| Item | Details |
+| --- | --- |
+| Base model | [unsloth/Meta-Llama-3.1-8B-bnb-4bit](https://huggingface.co/unsloth/Meta-Llama-3.1-8B-bnb-4bit) |
+| Fine-tuning | LoRA / supervised fine-tuning |
+| Base-model loading | 4-bit bitsandbytes checkpoint |
+| Training dataset | [Turkish QA Multi-Dialog](https://huggingface.co/datasets/sixfingerdev/turkish-qa-multi-dialog-dataset) |
+| Published weights | Adapter weights, **not** the full base model |
 
----
+The base model identifier is taken from this repository's `adapter_config.json`.
 
 ## Installation
 
-Install required dependencies:
-
-```!pip install transformers accelerate bitsandbytes peft```
-
-Ensure you have a GPU with sufficient VRAM for 4-bit inference.  
-
----
-
-## Loading the Model
-
-1. **Load the Base Model**
-
-```from transformers import AutoTokenizer, AutoModelForCausalLM```
-
-```base_model = AutoModelForCausalLM.from_pretrained(
-    "unsloth/llama-3.1-8b-bnb-4bit",
-    device_map="auto"
-)
+```bash
+pip install -U transformers accelerate peft bitsandbytes torch
 ```
 
-2. **Load the Adapter**
-
-'from peft import PeftModel'
-
-'model = PeftModel.from_pretrained('
-'    base_model,'
-'    "sixfingerdev/SixFinger-8B"'
-')'
+A compatible GPU and enough memory to load the quantized base model are required.
 
-3. **Load the Tokenizer**
+## Loading and generating text
 
-'tokenizer = AutoTokenizer.from_pretrained("unsloth/llama-3.1-8b-bnb-4bit")'
-
----
-
-## Example Usage
-
-Generate text using the adapter:
 ```python
-from transformers import AutoTokenizer, AutoModelForCausalLM
-from peft import PeftModel
 import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
 
-# Base model
-base_model = AutoModelForCausalLM.from_pretrained(
-    "unsloth/llama-3.1-8b-bnb-4bit",
-    device_map="auto"
-)
+base_id = "unsloth/meta-llama-3.1-8b-bnb-4bit"
+adapter_id = "sixfingerdev/SixFinger-8B"
 
-# LoRA adapter
-model = PeftModel.from_pretrained(base_model, "sixfingerdev/SixFinger-8B")
+tokenizer = AutoTokenizer.from_pretrained(base_id)
+base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
+model = PeftModel.from_pretrained(base, adapter_id)
+model.eval()
 
-# Tokenizer
-tokenizer = AutoTokenizer.from_pretrained("unsloth/llama-3.1-8b-bnb-4bit")
+prompt = "Soru: Yapay zekâ nedir?\nCevap:"
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
 
-# Örnek text generation
-prompt = "Soru: Yapay zeka nedir?\nCevap:"
-inputs = tokenizer(prompt, return_tensors="pt")
-with torch.no_grad():
-    outputs = model.generate(**inputs, max_new_tokens=50, do_sample=True, temperature=0.7)
+with torch.inference_mode():
+    output_ids = model.generate(
+        **inputs,
+        max_new_tokens=128,
+        do_sample=False,
+        pad_token_id=tokenizer.eos_token_id,
+    )
 
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+new_tokens = output_ids[0, inputs["input_ids"].shape[-1]:]
+print(tokenizer.decode(new_tokens, skip_special_tokens=True))
 ```
----
-
-## Notes
 
-- The adapter does **not** modify the base model; it only applies LoRA weights on top.  
-- 4-bit quantization significantly reduces VRAM usage. Ensure your GPU supports **bitsandbytes 4-bit operations**.  
-- You can merge the adapter into the base model for easier deployment if needed.  
+This is a loading example, not a verified quality benchmark. Exact outputs depend on the prompt, decoding settings, and installed library versions.
 
----
+## Evaluation and limitations
 
-## References
+- **Experimental checkpoint:** no independently reproducible benchmark results have been published for this adapter.
+- The repository is **not** an 8B model trained from scratch; its behavior depends on the upstream model.
+- Generated information may be incorrect. Evaluate on your own use case before deployment.
+- Data quality and provenance should be reviewed before production use.
 
-- [PEFT (Parameter-Efficient Fine-Tuning)](https://huggingface.co/docs/peft/index)  
-- [Transformers 4-bit Quantization](https://huggingface.co/docs/transformers/main/en/main_classes/quantization)  
-
----
+## License and attribution
 
-## License
+The adapter is published with the repository's **Apache-2.0** license metadata. **The underlying Llama 3.1 base model is separately governed by the [Llama 3.1 Community License](https://huggingface.co/meta-llama/Llama-3.1-8B/blob/main/LICENSE).** Combining or redistributing model weights requires complying with the applicable upstream terms.
 
-The adapter and its usage are provided under the terms specified in the repository.  
-Ensure compliance with the **base model license** (Meta’s LLaMA).
+See [PEFT documentation](https://huggingface.co/docs/peft/index) for loading and merging adapters.