Anis

Developed by: At-Tawheed · Attawheed AI Lab (ATTLAB)
Base model: unsloth/qwen2.5-7b-unsloth-bnb-4bit
Parameters: 8B · Tensor type: BF16 · License: Apache 2.0

This Qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

---

About Anis

Anis is an 8B parameter language model fine-tuned from Qwen2.5-7B using Supervised Fine-Tuning (SFT). It is the first stage in ATTLAB's open-source RLHF alignment pipeline, trained on At-Tawheed/Anis-RLHF — a curated 57.9 GB dataset of 33 instruction, preference, math, code, and multilingual subsets.

unsloth/qwen2.5-7b-unsloth-bnb-4bit  (base)
    └── Anis                           (SFT  ← this model)
            └── attlab-7b-dpo-v1       (DPO)

System prompt:

You are ATTLAB, a helpful, harmless, and honest AI assistant developed by the ATTLAB team.

---

Training Data — Anis-RLHF (57.9 GB · 33 subsets)

Category	Subsets
Instruction / Chat	openhermes_2_5, slim_orca, openorca_full, ultrachat_200k, smoltalk_1m, lmsys_chat_1m, tulu3_sft_mixture, oasst1_top_ranked
Preference / DPO	ultrafeedback, ultrafeedback_binarized, hh_rlhf_full, capybara_dpo_7k, helpsteer2
Math / Reasoning	metamath_qa, numina_math_cot, openmath_instruct2, magpie_reasoning_250k
Code	opencode_instruct_5m, codefeedback_66k, evol_codealpaca_110k, magicoder_oss_75k
Synthetic	magpie_llama3_1m, magpie_llama31_1m, magpie_llama33_1m, magpie_qwen25_1m
Knowledge	wikipedia_en, gutenberg_books, stackexchange_qa, fineweb_edu
Multilingual	aya_multilingual, wikipedia_yoruba
WizardLM	wizardlm_evol_v2

---

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "At-Tawheed/Anis"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are ATTLAB, a helpful, harmless, and honest AI assistant developed by the ATTLAB team."},
    {"role": "user",   "content": "What is reinforcement learning from human feedback?"}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        input_ids,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
    )

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

With Unsloth

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="At-Tawheed/Anis",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

---

Limitations

• SFT only: Anis is not fully aligned. For preference-optimized outputs use the DPO variant (attlab-7b-dpo-v1).
• Hallucination: May produce factually incorrect outputs — do not use as a sole source of truth.
• Bias: Training data is sourced from the internet and inherits its biases.

---

Citation

@misc{anis2025,
  author       = {Ibraheem, Olushola Taoheed},
  title        = {Anis: A Supervised Fine-Tuned Language Model},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/At-Tawheed/Anis}},
  note         = {Attawheed AI Lab (ATTLAB). Fine-tuned from Qwen2.5-7B with Unsloth and TRL.}
}

---

ATTLAB · Hugging Face · GitHub · Dataset