base_model: unsloth/gpt-oss-20b-unsloth-bnb-4bit tags: - text-generation-inference - transformers - unsloth - gpt_oss - agentic - tool-use - kilo-code license: apache-2.0 language: - en
Kilo Code Tool-Specialized Fine-Tuned Model Developed by: hybridfree
License: Apache-2.0
Base model: unsloth/gpt-oss-20b-unsloth-bnb-4bit
Overview This model is a task-oriented, agentic fine-tune of GPT-OSS-20B, trained on a carefully curated dataset focused exclusively on Kilo Code tool and function usage.
Unlike general instruction-tuned models, this model was optimized for:
Accurate tool invocation
Correct function argument construction
Strict adherence to Kilo Code execution flow
Reduced hallucinated logs and fake task completion
Long-horizon task consistency
The training data emphasizes realistic tool traces, partial executions, retries, and failure states, ensuring the model learns how to operate rather than merely how to explain.
Training Details Fine-tuned using Unsloth for high-efficiency training
Leveraged Hugging Face TRL for supervised fine-tuning
Achieved ~2× faster training compared to standard pipelines
Base model quantized to 4-bit (bnb) for efficient inference without sacrificing task reliability
Intended Use This model is designed for:
Kilo Code–based agent systems
Tool-driven code execution workflows
Autonomous or semi-autonomous coding agents
Verifier-compatible execution pipelines
It is not optimized for casual chat or creative writing. Its primary objective is execution correctness over conversational polish.
Notes For best results, pair this model with:
A strict tool-execution environment
A verifier or auditor pass to enforce artifact-based completion
This ensures maximum reliability and eliminates premature or fabricated task completion.
Training Framework This model was trained using Unsloth and Hugging Face’s TRL library.
- Downloads last month
- 32