livebylord/Qweblethos-v1-GGUF-EXPEREMENTIAL

Qweblethos v1 — Qwen3.8-27B Fable Agentic Fine-Tune (Q4KM GGUF) TEST VERSION!

Qweblethos v1 is a Q4_K_M GGUF of Qwen/Qwen3.8-27B, fine-tuned for coding-agent behavior, multi-step tool use, debugging, and verification-oriented workflows.

The name combines Qwen + Fable + Mythos. That is the project identity, not a provenance claim: this v1 checkpoint was trained on Fable 5 traces only; no Mythos-generated samples were used. A future version may add a separately audited Mythos component.

What this model is

• —Base model: `Qwen/Qwen3.8-27B`, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• —Fine-tuning method: LoRA, rank 32, alpha 32
• —Trainable parameters: 159,383,552 (about 0.58% of 27,516,112,112 parameters)
• —Training context length: 16,384 tokens
• —Training examples: 6,471
• —Evaluation examples used during the run: 128
• —Training: one epoch, 809 optimizer steps, one NVIDIA B200, about 2 hours
• —Export: LoRA merged into the BF16 base, then quantized to GGUF v3 Q4_K_M
• —Runtime target: llama.cpp-compatible applications
• —Vision and MTP: not included in this text-only export
• —Experemental finetune. The new v2 is coming!

This is response/trajectory distillation through supervised fine-tuning on synthetic agent traces. It is not logit-level distillation and it is not a reproduction of the teacher model.

Training data

The final training mixture contains:

The preparation pipeline removed duplicates, filtered invalid or overlong samples, excluded service/meta noise, converted tool calls to the native Qwen3.8 format, supervised assistant tokens only, and kept train/evaluation task clusters disjoint.

Evaluation

The same 128 held-out examples were evaluated with the base model and with the trained adapter enabled:

This is a 36.76% relative loss reduction on the prepared held-out distribution. It shows that the adapter learned that distribution. It does not prove an improvement on independent coding benchmarks such as SWE-bench, LiveCodeBench, or Terminal-Bench. No independent benchmark score is claimed for v1.

Downloaded file

The export was produced with llama.cpp commit 0d9ceae1e38291035605613ab41a8f5e693d6fcd and passed a CPU smoke test by generating the exact requested text hello world.

Running with llama.cpp

llama-cli \
  -m Qweblethos-v1-Q4_K_M.gguf \
  -c 16384 \
  --jinja \
  -p "Fix this bug, explain the cause, and add focused tests."

OpenAI-compatible local server:

llama-server \
  -m Qweblethos-v1-Q4_K_M.gguf \
  -c 16384 \
  --jinja \
  --host 127.0.0.1 \
  --port 8080

Adjust GPU offload (-ngl) to match your hardware. The model can run with CPU/RAM offload, but a 27B model remains heavy: 24 GB of system RAM is marginal, while 32 GB or more is the practical target. Small GPUs can offload only part of the model.

Intended use

Qweblethos v1 is intended for local experimentation with:

• —coding-agent workflows;
• —debugging and test-driven fixes;
• —multi-step tool use;
• —repository inspection and verification-oriented responses;
• —llama.cpp and OpenAI-compatible local serving.

Limitations

• —The training distribution is narrow and heavily focused on coding-agent traces.
• —General chat, multilingual quality, vision, and unrelated reasoning domains may regress relative to the base model.
• —Tool-call formatting still depends on the runtime and chat-template implementation.
• —The model may hallucinate commands, files, test results, or tool outputs. Treat generated actions as untrusted and review them before execution.
• —This release has a smoke test and held-out loss comparison, not a comprehensive independent benchmark suite.
• —Quantization can reduce quality compared with the merged BF16 checkpoint.

License and attribution

This release combines components with different upstream terms. The Qwen3.8-27B base is Apache-2.0; the selected Fable datasets include AGPL-3.0 and CC BY 4.0 material. See `LICENSE_NOTICE.md` before redistribution or commercial use.

Reproducibility summary

• —Base revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• —Dataset preparation revisions: lordx64 9df06dd, Crownelius/Glint b9f20b5
• —LoRA: rank 32, alpha 32, dropout 0
• —Learning rate: 8e-5
• —Effective batch size: 8
• —Max sequence length: 16,384
• —Final train loss: 0.51743
• —Final evaluation loss: 0.46328

Qweblethos is an independent community fine-tune and is not affiliated with Qwen, Alibaba, Anthropic, Fable, Mythos, Glint Research, or the dataset authors.