Xx-Vexento-xX/roast-bot-qwen-1b-epoch2

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The Xx-Vexento-xX/roast-bot-qwen-1b-epoch2 is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned by Xx-Vexento-xX on a unique dataset reflecting Sarvin's WhatsApp texting style. This specific checkpoint, taken at epoch 2, is optimized to generate original, synthesized responses rather than recalling verbatim training examples. It offers a balance between generalization and coherence, making it suitable for applications requiring creative text generation in a specific conversational style.

Loading preview...

Roast Bot - Qwen 1.5B (Epoch 2 Checkpoint)

This model is a fine-tuned version of Qwen2.5-1.5B-Instruct, developed by Xx-Vexento-xX, specifically trained on Sarvin's WhatsApp texting style. It utilizes LoRA (r=32) for efficient adaptation and represents an earlier checkpoint (epoch 2 of 5) from its training run.

Key Differentiators & Capabilities

  • Original Generation: Unlike its fully-trained counterpart (epoch 5), this epoch 2 checkpoint prioritizes generating novel responses, even on inputs similar to the training data, rather than verbatim recall.
  • Generalization: With an evaluation loss of 0.661 at epoch 2, it demonstrates better generalization compared to the epoch 5 model, which showed signs of overfitting (eval loss 0.989).
  • Specific Style Emulation: The model is designed to emulate a distinct WhatsApp texting style, based on 90 hand-written conversation pairs.

When to Use This Checkpoint

  • Genuine Generation: Choose this model if your primary requirement is for genuinely synthesized responses from the input, accepting a slight trade-off in surface polish or occasional coherence compared to memorized outputs.
  • Avoidance of Overfitting: It's ideal for scenarios where avoiding the exact recall of training examples is crucial, ensuring more dynamic and less predictable interactions.

Training Details

  • Base Model: unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit
  • Methodology: LoRA (r=32, alpha=64, dropout=0.05) via Unsloth.
  • Dataset: 90 hand-written conversation pairs (80 train / 10 eval split).
  • Checkpoint: Epoch 2 of 5, with an evaluation loss of 0.661.

Finetunes
1 models