Shekswess/trlm-135m

Tiny Reasoning Language Model (trlm-135)

Table of Contents

1. Model Summary
2. Post-Training Pipeline
3. How to use
4. Training
5. Evaluation
6. Limitations
7. Acknowledgements
8. License

---

Model Summary

The Tiny Reasoning Language Model (trlm-135) is a 135M parameter research prototype designed to study how small models can learn step-by-step reasoning. It was built on top of SmolLM2-135M-Instruct and fine-tuned through a 3-stage pipeline:

• Stage 1 SFT: general instruction tuning (non-reasoning).
• Stage 2 SFT: reasoning traces with <think> tags.
• Stage 3 DPO: preference alignment for reasoning style.

The code for everything can be found here

---

Post-Training Pipeline

---

How to use

pip install -U transformers accelerate

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Shekswess/trlm-135m"
device = "cuda"  # or "cpu"

# Load tokenizer & model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
).to(device)

# Example prompt
prompt = "Give me a brief explanation of gravity in simple terms."
messages = [
    {"role": "user", "content": prompt}
]

# Apply chat template
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

For reasoning-heavy tasks, set temperature=0.6 and top_p=0.95.

---

Training

Model

• Architecture: Decoder-only transformer (SmolLM2 backbone which infact is Llama 3 based model).
• Parameters: ~135M.
• Precision: mix-precision (bfloat16) during training.

Software & Hardware

• Training Frameworks: PyTorch (ROCm), Hugging Face Transformers & TRL.
• Hardware: AMD MI300X (192GB VRAM, 224GB RAM).

Special thanks to @HotAisle

Training Stages

1. Stage 1 – SFT (non-reasoning)
   • ~58k samples, everyday conversations & instruction following.
2. Stage 2 – SFT (reasoning)
   • ~78k samples with <think> segments.
3. Stage 3 – DPO (alignment)
   • ~50k preference pairs (chosen vs. rejected reasoning traces).

---

Evaluation

Evaluation was done with lm-eval-harness:

Benchmark	Tiny Reasoning Language Model (trlm-135M)	SmolLM2-135M-Instruct	Improvements
ARC Challenge	40.61 (avg)	37.3 (avg)	+3.31
BBH	36.80 (3-shot)	28.2 (3-shot)	+8.6
BoolQ	62.17	–	N/A
GSM8K	2.59 (5-shot)	1.4 (5-shot)	+1.19
IFEval	35.49 (avg)	29.9 (avg)	+5.59
MMLU	34.95	29.3	+5.65
PIQA	64.91	66.3	–1.39
HellaSwag	–	40.9	N/A
MT-Bench	–	19.8	N/A

---

Limitations

• Not production-ready: hallucinations and logical errors are frequent.
• Small size: limited general knowledge and reasoning depth.
• English-only: multilingual capabilities not explored.

---

Acknowledgements

• @HotAisle for providing the compute resources to train all three stages on a awesome AMD MI300x setup.
• @mkurman88 for ideas, feedback and code samples.
• HuggingFaceTB team for SmolLM2-135M-Instruct model and the Smoltalk2 dataset collection.
• @scottgeng00 for the OLmO-3-Preference-Mix-Deltas dataset.
• @eliebakouchi for help with the tokenization.

---

License

Apache 2.0

---