trlm-stage-2-sft-final-2 is the Stage 2 post-training model for the Tiny Reasoning Language Model (trlm) project.
This stage focuses on reasoning tasks, fine-tuned on a curated dataset of 78,000 entries with reasoning tokens (<think>...</think>).

📖 Model Description

Base Model: Shekswess/trlm-stage-1-sft-final-2
Type: Causal Language Model (decoder-only transformer)
Stage: Post-training Stage 2 (SFT)
Objective: Equip the model with reasoning ability, multi-turn thought structuring, and explicit <think> chain-of-thought representations.

This stage teaches the model to analyze problems step-by-step, reason with intermediate thoughts, and provide structured answers.

🎯 Intended Uses & Limitations

Intended Uses

Reasoning-based question answering
Step-by-step logical explanations
Multi-turn reasoning with <think> traces
Precursor to preference optimization (Stage 3)

Limitations

May overfit on reasoning style and hallucinate <think> tokens in simple tasks
Still limited in knowledge scope (135M parameters)
Trained only on English datasets

📊 Training Data

This model was trained on the dataset:
👉 Shekswess/trlm-sft-stage-2-final-2

Dataset summary:

Entries: 78,000
Sources: 6 HuggingFaceTB/smoltalk2 subsets
Focus: Reasoning tasks with <think> annotations

Source Dataset	Entries	Percentage %
Llama_Nemotron_Post_Training_Dataset_reasoning_r1	40,200	51.5%
OpenThoughts3_1.2M	20,000	25.6%
multi_turn_reasoning_if_think	10,000	12.8%
aya_dataset_Qwen3_32B_think	5,000	6.4%
smoltalk_everyday_convs_reasoning_Qwen3_32B_think	2,000	2.6%
s1k_1.1_think	800	1.0%

⚙️ 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: 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-2-sft-final-2"

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

# Example inference with reasoning
messages = [
    {"role": "user", "content": "If a train travels 60 km in 1 hour and another 90 km in 1.5 hours, what is the average speed?"}
]

# 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))

📌 Next Steps

Stage 3: DPO / preference optimization for reasoning stability

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