FF_3.1 is a 2.02B parameter GPT-2 decoder-only language model trained from scratch with a multi-stage pipeline combining supervised fine-tuning, preference optimization, knowledge distillation, and instruction tuning.

Model Details


Architecture	GPT-2 decoder-only
Parameters	2.02B
Hidden size (d)	2048
Attention heads (h)	16
FFN size (ff)	8192
Layers (L)	38
Context length	2048
Tokenizer	GPT-2 BPE (vocab size: 50,257)
Precision	bfloat16

Training Pipeline

FF_3.1 was trained through a 5-stage pipeline:

Pretraining — 90B tokens on a large English corpus
SFT — 760K + 100K examples (OpenHermes-2.5 / NuminaMath / Eurus)
DPO — 38,863 preference pairs
Distillation v3 — 47K examples targeting MMLU + GSM8K + ARC benchmarks
LoRA v4b — 10K examples for instruction following refinement

Evaluation

Benchmark	Score
MMLU (5-shot)	27.94% (+3.94 pp vs FF_3 baseline of 24%)

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("francescofiamingo1/FF_3.1", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("francescofiamingo1/FF_3.1")

input_text = "Explain photosynthesis in simple terms."
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Known Limitations

Math reasoning is still weak — the model struggles with multi-step arithmetic and word problems
Instruction count following is imprecise — the model may not reliably follow constraints like "list exactly 5 items"

What's Next

FF_3.2 will focus on:

DPO with UltraFeedback dataset for improved preference alignment
Improved math dataset for stronger quantitative reasoning

License

Apache 2.0