README.md

---
language:
- en
license: apache-2.0
base_model: EleutherAI/gpt-j-6b
tags:
- gpt-j
- lora
- peft
- reasoning
- thinking
- sft
- qwen3-template
datasets:
- nohurry/Opus-4.6-Reasoning-3000x-filtered
- TeichAI/claude-4.5-opus-high-reasoning-250x
- Jackrong/Qwen3.5-reasoning-700x
pipeline_tag: text-generation
---

# GPT-J 6B — Thinking SFT · LoRA Adapter

Raw **LoRA adapter** (QLoRA, r=32, α=32) for
[EleutherAI/gpt-j-6b](https://huggingface.co/EleutherAI/gpt-j-6b), trained with
the Qwen3 chain-of-thought format.

The **fully merged float16 model** (ready to use without PEFT) is at  
👉 [ping98k/gpt-j-6b-thinking-sft](https://huggingface.co/ping98k/gpt-j-6b-thinking-sft)

---

## Adapter Details

| Property | Value |
|---|---|
| **Base model** | EleutherAI/gpt-j-6b |
| **LoRA rank** | 32 |
| **LoRA alpha** | 32 |
| **Target modules** | q_proj, k_proj, v_proj, out_proj, fc_in, fc_out |
| **Trainable parameters** | ~50 M |
| **Vocabulary additions** | `<|im_start|>`, `<|im_end|>`, `<think>`, `</think>` (50 404 total) |

---

## Usage with PEFT

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

base_model_id = "EleutherAI/gpt-j-6b"
adapter_id    = "ping98k/gpt-j-6b-thinking-sft-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)  # resized vocab

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)
model.resize_token_embeddings(len(tokenizer))
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()
```

> **Tip:** For most use cases, prefer the pre-merged model at
> [ping98k/gpt-j-6b-thinking-sft](https://huggingface.co/ping98k/gpt-j-6b-thinking-sft) —
> no PEFT dependency needed.

---

## License

Apache 2.0 (inherits from EleutherAI/gpt-j-6b).