ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 26, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint is a 3.1 billion parameter causal language model based on the Qwen2ForCausalLM architecture. It is a LoRA fine-tune of WeiboAI/VibeThinker-3B, specifically supervised fine-tuned on the private ermiaazarkhalili/Fable-5-Glint-Clean dataset. This model demonstrates a significant improvement in next-token accuracy on its training data compared to its base model, making it suitable for tasks aligned with its instruction-following fine-tuning.

Loading preview...

Model Overview

ermiaazarkhalili/VibeThinker-3B-SFT-Fable5-Glint is a 3.1 billion parameter language model built on the Qwen2ForCausalLM architecture. It is a LoRA (Low-Rank Adaptation) fine-tune of the WeiboAI/VibeThinker-3B base model, specifically trained using supervised fine-tuning (SFT) on a private dataset, ermiaazarkhalili/Fable-5-Glint-Clean.

Key Characteristics

  • Base Model: WeiboAI/VibeThinker-3B
  • Architecture: Qwen2ForCausalLM
  • Parameters: 3.1 billion
  • Fine-tuning Method: LoRA SFT via Unsloth and TRL, utilizing QLoRA (4-bit precision).
  • Training Data: ermiaazarkhalili/Fable-5-Glint-Clean (private instruction-following dataset).
  • Context Length: Trained with a maximum sequence length of 4096 tokens.

Performance Highlights

Evaluation on a held-out split of the training data shows notable improvements over the base model:

  • Top-1 Accuracy: Increased from 0.5309 (base) to 0.6719 (this model), a delta of +0.1411.
  • Top-5 Accuracy: Increased from 0.7563 (base) to 0.8947 (this model), a delta of +0.1384.

These metrics indicate that the fine-tuning process effectively adapted the model to the distribution of the Fable-5-Glint-Clean dataset.

Limitations

  • No external benchmark evaluations have been conducted; performance metrics are limited to training-loss observations and held-out accuracy on the fine-tuning dataset.
  • Inherits biases, knowledge cutoffs, and potential failure modes from its base model.
  • Behavior outside the specific instruction-following distribution of its fine-tuning dataset is untested.
  • LoRA adapters are merged, meaning the model cannot be easily reverted to its pre-fine-tuned state.

Finetunes
2 models