Part of the Parable series: small local LLMs fine-tuned on genuine agent
traces. This is the reasoning-style chat variant of Nanbeige4.2-3B (to our
knowledge the first published fine-tune of this base), tuned on real Claude
Fable 5 agent transcripts so its step-by-step reasoning voice carries into
local use.
Nanbeige support landed in llama.cpp master on 2026-07-27, so you
currently need llama.cpp built from master. As of 2026-07-29 no packaged release has
it yet: Homebrew's bottle (build 10150) still fails with
unknown model architecture: 'nanbeige' (verified), and Ollama / LM Studio
bundle older builds too. Build from source until the next releases land:
1llama-cli -m Parable-Nanbeige4.2-3B-Claude-Fable-5-Q4_K_M.gguf \2 -ngl 99 -c 4096\3 -p "Write a python function that reverses a string." --temp 0.6
Runtime status (2026-07-29, stated plainly): the merged weights generate
correctly (verified via transformers on GPU; sample outputs in the artifacts
repo under verify/), and these GGUFs converted and quantized cleanly with
num_loops=2 in the header. However, in our tests CPU-only inference of
this architecture hangs with current llama.cpp master builds; the
architecture is days old and its CPU path appears immature upstream. Prefer
GPU offload (-ngl 99), or use the
full-precision repo
with transformers until upstream stabilizes. We will update this note as
llama.cpp matures.
Output begins with a <think>...</think> reasoning block, then the answer.
If you are building on top of this model, parse and strip the think block
before showing text to end users.
The architecture runs its 22-layer stack twice (num_loops: 2, effective
depth 44). This is encoded in the GGUF and handled automatically; expect
roughly 2x the per-token compute of a typical 22-layer model of this width.
Method: QLoRA (nf4, r=16, alpha=32) on the standard 7-module target set, all 22 layers (154 adapted modules)
Data: 845 genuine Claude Fable 5 agent-session traces (chat variant: prose answers, <think> reasoning preserved) + 362 replay rows (OpenCoder educational_instruct + tulu-3 mix) to limit forgetting; 98% of rows retain their full assistant span at the 3,072-token training window
Schedule: 380 optimizer steps, effective batch 8, completion-only loss masking, cosine LR
Held-out trace loss: 2.333 (base) → 2.133 (tuned), same 225-trace held-out set never used in training
Evaluation
The number above is a held-out language-modeling loss on real agent traces:
it measures how much better the model fits the Fable-5 reasoning
distribution, and it is the honest headline for what this fine-tune does.
We do not claim benchmark gains over the base model. Nanbeige4.2-3B's model
card reports strong agentic results (SWE-bench Verified 63.6, Terminal-Bench
2.0 44.1, vendor self-reported); this variant trains the reasoning voice,
not tool-calling, and prose-only SFT should be expected to trade a little
benchmark sharpness for style fidelity. If you need maximum benchmark
performance on agentic harnesses, use the base; if you want its capability
with a transparent, well-structured reasoning trace in local use, use this.
Limitations
The base model's own benchmark figures are vendor self-reported and were
not independently reproduced here.
Training ran on a 4-bit quantized base (QLoRA); the F16 merge cannot exceed
4-bit-base quality.
Requires llama.cpp built from master (2026-07-27+). No packaged release
(brew, Ollama, LM Studio) loads this architecture yet; they will as their
bundled llama.cpp catches up.
Quantization
Quantized with llama.cpp llama-quantize from the F16 merge, num_loops=2
verified present in the GGUF header at conversion time.
Provenance & licensing
Fine-tuned from Nanbeige/Nanbeige4.2-3B (Apache-2.0). Training data:
Glint-Research/Fable-5-traces
(AGPL-3.0) and
Roman1111111/gpt5.5-terminal
(MIT). Because those traces originate from third-party assistants, the
providers' terms may apply to downstream training and distillation. If you
plan to build on this model commercially, confirm your use aligns with those
terms.
Citation
The recipe, evaluation methodology and failure analysis behind this model are
documented in the tech report:
Aglawe, A. (2026). Agent-Trace Fine-Tuning of Small Language Models under
Constrained Compute. Zenodo. doi:10.5281/zenodo.21676407
bibtex
1@misc{aglawe2026agenttrace,
2 author = {Aglawe, Ankit},
3 title = {Agent-Trace Fine-Tuning of Small Language Models under Constrained Compute},
4 year = {2026},
5 publisher = {Zenodo},
6 doi = {10.5281/zenodo.21676407},
7 url = {https://doi.org/10.5281/zenodo.21676407}
8}
Acknowledgements
The Nanbeige team for the base model; Glint-Research and Roman1111111 for the
trace datasets; empero-ai for the recipe this series iterates on.
