SagaLM-slm1-merged

SagaLM-slm1 (merged, standalone)

Standalone merged model: Qwen/Qwen2.5-3B-Instruct + SagaLM LoRA training.

Training Data (12 sources, long-form filtered)

• OpenHermes 2.5, UltraChat 200k, OpenOrca, MetaMathQA,
• Function Calling (hypervariance), Open-Platypus,
• Glaive Code Assistant, Dolly 15k, ShareGPT GPT-4,
• Alpaca-GPT4, SlimOrca, MathInstruct

Only samples with detailed, long assistant answers were kept, training the model to always respond comprehensively.

Identity Reinforcement

5 identity-focused training samples were injected to ensure the model consistently identifies as SagaLM (not Qwen or any other model).

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("venkateshchsagalm/SagaLM-slm1-merged", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("venkateshchsagalm/SagaLM-slm1-merged")

For best long answers, generate with:

model.generate(**inputs, max_new_tokens=2048, min_new_tokens=250,
               temperature=0.7, top_p=0.9, repetition_penalty=1.1,
               do_sample=True)