---
library_name: transformers
license: apache-2.0
base_model: Shekswess/trlm-stage-2-sft-final-2
tags:
- trl
- dpo
- preference-alignment
- reasoning
- generated_from_trainer
model-index:
- name: trlm-stage-3-dpo-final-2
  results: []
---

![image/png](https://github.com/user-attachments/assets/5f453496-8180-4cf4-94da-26ebbe1159d4)


# 🧠 trlm-stage-3-dpo-final-2

`trlm-stage-3-dpo-final-2` is the **Stage 3** post-training model for the **Tiny Reasoning Language Model (trlm)** project.  
This stage focuses on **preference alignment** using **Direct Preference Optimization (DPO)** with 50k preference pairs.

---

## 📖 Model Description

- **Base Model**: [Shekswess/trlm-stage-2-sft-final-2](https://huggingface.co/Shekswess/trlm-stage-2-sft-final-2)  
- **Type**: Causal Language Model (decoder-only transformer)  
- **Stage**: Post-training **Stage 3 (DPO)**  
- **Objective**: Align model outputs with human-preferred reasoning and answers by contrasting **chosen** vs **rejected** completions.  

This stage improves the model’s **alignment**, **coherence**, and **reasoning stability**.

---

## 🎯 Intended Uses & Limitations

### Intended Uses
- Aligned reasoning assistant with structured `<think>` traces  
- Multi-turn reasoning with preference-optimized outputs  
- Safer, more useful responses for reasoning tasks  

### Limitations
- Trained only on preference data → may inherit biases from source datasets  
- Limited parameter count (135M) restricts knowledge breadth  
- Still prone to hallucinations under complex reasoning chains  

---

## 📊 Training Data

This model was trained on the dataset:  
👉 [**Shekswess/trlm-dpo-stage-3-final-2**](https://huggingface.co/datasets/Shekswess/trlm-dpo-stage-3-final-2)

**Dataset summary**:
- **Entries**: 50,000 preference pairs  
- **Source**: `scottgeng00/olmo-3-preference-mix-deltas_reasoning-yolo_scottmix-DECON-chfiltered`  
- **Focus**: Preference alignment with **chosen vs rejected responses**  

| Source Dataset | Split | Entries | % |
|----------------|-------|---------|---|
| scottgeng00/olmo-3-preference-mix-deltas_reasoning-yolo_scottmix-DECON-chfiltered | train | 50,000 | 100% |

---

## ⚙️ Training Procedure

### Training Hyperparameters
- **Learning rate**: 1e-5  
- **Train batch size**: 32  
- **Eval batch size**: 8  
- **Gradient accumulation steps**: 4  
- **Total effective batch size**: 128  
- **Optimizer**: AdamW (betas=(0.9, 0.999), eps=1e-08)  
- **LR Scheduler**: Cosine with minimum LR + warmup ratio 0.1  
- **Epochs**: 1  
- **Seed**: 42  

### Framework Versions
- **Transformers**: 4.56.2  
- **PyTorch**: 2.7.1+rocm7.0.0.git698b58a9  
- **Datasets**: 4.0.0  
- **Tokenizers**: 0.22.1  

---

## 🚀 Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Shekswess/trlm-stage-3-dpo-final-2"

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

# Example inference with preference-aligned reasoning
messages = [
    {"role": "user", "content": "Explain why the sky is blue in simple terms."}
]

# Apply chat template
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt")

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

--- 

Part of the Tiny Reasoning Language Model (trlm) post-training pipeline.