trlm-stage-1-sft-final-2/README.md

---
library_name: transformers
license: apache-2.0
base_model: HuggingFaceTB/SmolLM2-135M-Instruct
tags:
- trl
- sft
- generated_from_trainer
model-index:
- name: trlm-stage-1-sft-final-2
  results: []
---


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

# 🧠 trlm-stage-1-sft-final-2

`trlm-stage-1-sft-final-2` is the **Stage 1** post-training model for the **Tiny Reasoning Language Model (trlm)** project.  
This stage focuses on **everyday conversations** and **general instruction following**, fine-tuned on a curated dataset of **58,000 entries**.

---

## 📖 Model Description

- **Base Model**: [HuggingFaceTB/SmolLM2-135M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct)  
- **Type**: Causal Language Model (decoder-only transformer)  
- **Stage**: Post-training **Stage 1 (SFT)**  
- **Objective**: Build a solid foundation in **instruction-following** and **dialogue coherence** before advancing to reasoning-specific training.  

This stage teaches the model to **follow instructions, rewrite, summarize, and hold conversations** without reasoning tokens.

---

## 🎯 Intended Uses & Limitations

### Intended Uses
- Everyday conversation assistants  
- Instruction-following tasks (summarization, rewriting, simple dialogue)  
- Precursor foundation for reasoning post-training (Stage 2+)  

### Limitations
- Not optimized for reasoning (handled in later stages)  
- May struggle with multi-step logical or mathematical problems  
- Trained only on English datasets  

---

## 📊 Training Data

This model was trained on the dataset:  
👉 [**Shekswess/trlm-sft-stage-1-final**](https://huggingface.co/datasets/Shekswess/trlm-sft-stage-1-final)

**Dataset summary**:
- **Entries**: 58,000  
- **Sources**: 7 HuggingFaceTB/smoltalk2 subsets  
- **Focus**: Non-reasoning conversations and instruction-following  

| Source Dataset | Entries | Percentage % |
|----------------|---------|---|
| smoltalk_smollm3_smol_magpie_ultra_no_think | 33,500 | 57.8% |
| smoltalk_smollm3_smol_summarize_no_think | 7,500 | 12.9% |
| smoltalk_smollm3_smol_rewrite_no_think | 7,500 | 12.9% |
| smoltalk_smollm3_systemchats_30k_no_think | 2,500 | 4.3% |
| smoltalk_smollm3_explore_instruct_rewriting_no_think | 2,500 | 4.3% |
| tulu_3_sft_personas_instruction_following_no_think | 2,500 | 4.3% |
| smoltalk_smollm3_everyday_conversations_no_think | 2,000 | 3.4% |

---

## ⚙️ Training Procedure

### Training Hyperparameters
- **Learning rate**: 3e-4  
- **Train batch size**: 32  
- **Eval batch size**: 8  
- **Gradient accumulation steps**: 4  
- **Total effective batch size**: 128  
- **Optimizer**: AdamW (betas=(0.9, 0.99), eps=1e-08)  
- **LR Scheduler**: Cosine with warmup ratio 0.1  
- **Epochs**: 2  
- **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-1-sft-final-2"

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

# Example inference
inputs = tokenizer("Write a short daily affirmation:", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## 📌 Next Steps

- **Stage 2**: Supervised fine-tuning with reasoning-focused data
- **Stage 3**: DPO / preference optimization for reasoning stability

---

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