🤖 Tiny OpenHermes — LoRA Fine-Tuned on OpenHermes-2.5
Fine-tuned TinyLlama/TinyLlama-1.1B-Chat-v1.0 on the teknium/OpenHermes-2.5 dataset using LoRA (rank 32) via TRL's SFTTrainer on Kaggle Dual T4 GPU.
LoRA rank 32 LoRA alpha 64 Epochs 1 Peak LR 2e-4 Effective batch 64 (4/GPU × 2 GPUs × 8 accum) Precision float16 Hardware Kaggle Dual T4 (2 × 16 GiB)
⚠️ Limitations English-primary (OpenHermes-2.5 is predominantly English)
May hallucinate facts — verify important claims
1.1 B parameter model: complex multi-step reasoning can fail
Not RLHF-aligned for safety beyond TinyLlama's base alignment
Benchmark Results
The model was evaluated using standard NLP benchmarks via the Language Model Evaluation Harness. It demonstrates moderate baseline capabilities in everyday physical reasoning but requires improvement in complex scientific knowledge and multi-step reasoning.
| Benchmark | Tasks (Samples) | Metric | Raw Score (acc) | Normalized Score (acc_norm) |
|---|---|---|---|---|
| PIQA (Physical Commonsense) | 1,838 | Accuracy | 72.58% | 72.03% |
| HellaSwag (Commonsense Reasoning) | 10,042 | Accuracy | 44.69% | 59.20% |
| ARC-Challenge (Advanced Science) | 1,172 | Accuracy | 25.43% | 29.69% |
| MMLU (mathemaatics) | 1531 | Accuracy | 26.13% | 26.13% |
🚀 Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"Havoc999/tiny-openhermes", torch_dtype=torch.float16
).cuda()
tok = AutoTokenizer.from_pretrained("Havoc999/tiny-openhermes")
prompt = "<|user|>\nExplain gravity simply.</s>\n<|assistant|>\n"
ids = tok(prompt, return_tensors="pt").input_ids.cuda()
out = model.generate(ids, max_new_tokens=200, temperature=0.7, do_sample=True)
print(tok.decode(out[0, ids.shape:], skip_special_tokens=True))