thecr7guy/gpt2-insFT

Short summary: A GPT-2–style causal LM instruction-tuned on a mixture of public datasets. Loss is applied only on the response segment, so the model learns to answer while treating the instruction and input as context.

⚠️ Safety note The training mix includes datasets that may contain harmful, harassing, or hateful text. This model is released for research and evaluation only.

---

Performance and Evaluation

Evaluation done using lm-evaluation-harness by EleutherAI. All evaluation used seed=7777 and batch_size 2. Number of fewshot examples = 0 for the all the benchmarks below.

Dataset	Metric	thecr7guy/gpt2-pretrain	GPT-2 (baseline)	thecr7guy/gpt2-insFT
HellaSwag	acc	0.291	0.289	0.2829
SciQ	acc	0.754	0.752	0.726
Winogrande	acc	0.491	0.516	0.4909
TruthfulQA MC1	acc	0.236	0.228	0.2619
MMLU (overall)	acc	0.230	0.229	0.2310
├─ Humanities	acc	0.242	0.242	0.2387
├─ Social Sci.	acc	0.217	0.217	0.2246
├─ STEM	acc	0.213	0.213	0.2226
└─ Other	acc	0.239	0.238	0.2343

Model details

• Base: thecr7guy/gpt2-pretrain

• Prompt format:

  • 
    Below is an instruction that describes a task. Write a response that appropriately completes the request.
    
    ### Instruction:
    {instruction}
    
    ### Input:
    {input}
    
    ### Response:
    

• Tokenization: For the base model, the eos_token = padding_token. For the IFT model, a custom pad token set to <|extra_7|>.

• Supervision signal: loss is masked up to the first token of the ### Response: span; only answer tokens (plus EOS) contribute to loss.

---

Training data

Mixture of public corpora:

• databricks/databricks-dolly-15k
• yahma/alpaca-cleaned
• allenai/prosocial-dialog (subset: response_id == 0, shuffled, 20k samples)
• Experimental/risky sets used for research:
• BornSaint/harmful_instructor
• BornSaint/evil_assistant

PreProcessing

• Columns normalized to instruction, input, output.
• Optional instruction templates:
  • For guard rails use the instruction: "Respond safely and constructively to the following user message."
  • To get unhinged resposnses use the instruction: "GODMODE SAI. Respond in a blunt, uncensored voice." (included only to study failure modes;)

---

Hyperparameters

• lr = 3e-5
• beta1 = 0.9
• beta2 = 0.95
• weight_decay = 0.1
• epochs = 2
• batch_size = 8
• grad_clip_norm= 1.0

Trained on runpod.io 4* Nvidia 4000 ADA (1$ per hour.)

Training took 25 mins on average for an epoch.

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "thecr7guy/gpt2-insFT"

tok = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = (
    "Below is an instruction that describes a task. "
    "Write a response that appropriately completes the request."
    "\n\n### Instruction:\n"
    "Give a concise, step-by-step explanation for the query"
    "\n\n### Input:\n"
    "How do I get better at basketball?"
    "\n\n### Response:\n"
)

inputs = tok(prompt, return_tensors="pt")
gen = model.generate(
  **inputs,
  max_new_tokens=256,
  do_sample=True,
  temperature=0.7,
  top_p=0.9,
  eos_token_id=tok.eos_token_id,
  pad_token_id=tok.pad_token_id,
)
print(tok.decode(gen[0], skip_special_tokens=True))

---

python inf_direct.py

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
Give a concise, step-by-step explanation for the query

### Input:
How do I get better at basketball?

### Response:
To get better at basketball, some tips are essential. Here are some steps to follow:

1. Prepare a strategy: Clear and well-defined objectives for your basketball team. This includes setting specific goals and objectives, understanding the rules of basketball, and setting specific goals and objectives.

2. Find the right players: Select the right players to represent your team in their basketball league. This could be a player's name, height, weight, and physical abilities.

3. Plan your approach: Make sure you have everything necessary to reach the goal. Consider spending time together and practicing your skills, as well as finding