Hikari Bonsai 27B Nimble LoRA — PQ2_0 GGUF

This repository provides Hikari-LoRA-fixed-scale-PQ2_0.gguf, a standalone GGUF artefact of a 27B model adapted by the publisher through LoRA training for Nimble. The name acknowledges Hikari07jp with respect and gratitude for their work. This derivative is published by zapabobouj; it is neither an official Hikari release nor an indication of Hikari's endorsement.

The account of training and the incorporation of LoRA updates into the GGUF is supported by cross-referenced records of the final training run, adapter checkpoint and export. No external LoRA adapter was specified during the recorded evaluation. The established details and the remaining uncertainties are set out below.

The model is intended for experimental and local use. The reported results measure agreement with a fixed set of recorded reference labels. They do not establish accuracy against independently adjudicated answers or demonstrate general capability.

Distributed Artefact

A PrismL build of llama.cpp supporting PQ2_0 is required. Compatibility and memory requirements have not been established for other builds or GPUs.

Provenance and LoRA Training Method

The publisher's Google Drive records were cross-referenced: FINAL_TRAIN_RESULT.json, the final adapter's adapter_config.json and COMPLETE.json, and Hikari-LoRA-fixed-scale-PQ2_0.receipt.json. The GGUF SHA-256 in the export receipt agrees with that of the distributed file. The adapter weights and configuration hashes also agree with the final checkpoint manifest. This account rests upon those retained records; training and conversion were not repeated in preparing the model card.

The source was the PQ2_0 model from Hikari's Ternary-Bonsai-2-27B-Abliterated-GGUF. The recorded source revision is e7f6daf95ab820ef8de7d8f5e883d95d546ab02c; the source GGUF SHA-256 is 41a362f422b70a8c2dc74a3cc14447ad0ea702c440f0dbe41dc1796da7b7e342.

Weight Representation and Target Modules

Training used a candidate model reconstructed from the Hikari GGUF in Hugging Face format. LoRA was applied to 96 auxiliary projection modules within the linear attention mechanism: linear_attn.in_proj_a and linear_attn.in_proj_b. Their source tensors were stored as BF16 in the original GGUF. The training record states that these target base weights were loaded as torch.float32.

The remaining base linear layers used an NF4 approximation of the reconstructed Hikari weights, with an unquantised lm_head. This was an experiment in LoRA adaptation of selected modules using an NF4 approximation of the base model. It did not train directly through native PQ2 operations.

The final record sets one_pass_of_usable_fold_complete=true and final_train_pass_complete=true. It also retains full_training_complete=false and full_fold_epoch_complete=false. Accordingly, the supported completion claim is confined to one pass through the 2,676 usable examples. During final training, the record reports final_test_opened=false and sealed_eval_jsonl_opened=false; the development subset contained no examples.

Incorporation of LoRA Updates into the GGUF

The final LoRA updates were incorporated into the target BF16 tensors of the original Hikari GGUF and written in their original GGUF type. The export receipt records updates to 96 of the 851 tensors, all BF16, with the remaining 755 source tensors retained. No PQ2 tensor was updated, no packed code bits were changed, and the original scale bytes were preserved. The distributed file therefore contains the incorporated LoRA updates and requires no separate adapter.

The exporter also supports a PQ2 overlay procedure which selects the nearest of four codes while fixing the original signed FP16 scales. For this particular artefact, however, the updates were confined to BF16 layers. The release is not described as a fresh PQ2 requantisation of the entire model.

The final adapter weights have SHA-256 4ecd41bd6193f82a8f249895671876b9d2b9f1e08c3d9b684fbd7b6cb287bdf0; the configuration has SHA-256 070382f337423e550281f550a14a8f263c3937ac9b9f8cda6eb633672e2a2908. These identifiers connect the training checkpoint records to the GGUF export receipt.

Limits to Reproducibility

The final result and configuration examined do not specify the applied learning rate, complete optimiser and scheduler settings, random seed, or full definition of the loss function. Example Nimble settings for another model have not been represented as the conditions of this 27B run. The training data permissions and complete execution environment remain unverified.

The final training record does not establish functional equivalence between the source GGUF and its reconstructed Hugging Face representation. It also records that the locally cached shard hashes were not rechecked during that run. The NF4 training base and the source GGUF into which updates were incorporated have different numerical representations; equivalence of behaviour or quality with the source model is therefore not assured. Independent adjudication of the 324 evaluation references and checks for overlap with the training data remain outstanding.

Intended Use and Constraints

Intended uses include text generation in a compatible local inference environment, experiments in candidate assessment with Nimble, and comparative evaluation against retained references. Model outputs are insufficient on their own to determine consequential medical, legal, financial, recruitment or public-service decisions. Access to external tools, confidential information and authorisation for consequential actions should be governed outside the model.

Evaluation

On 4 October 2026, the same 324 recorded Nimble evaluation cases were paired by identifier and their statistics recalculated retrospectively. Inference was not repeated in preparing this card. The quantities below are rates of agreement with recorded reference labels, rather than independently established answer accuracy.

Among the paired cases, seven agreed with the reference only under the comparator and 24 agreed only under the Nimble GGUF. The exact two-sided McNemar test gives a nominal p-value of 0.0033269 (b=7, c=24). The continuity-corrected McNemar statistic is χ²(1)=8.258, with p=0.004057.

The paired difference is +5.25 percentage points. The 95% confidence interval obtained by cluster bootstrapping at the family level is +2.16 to +8.33 percentage points.

Probability-Based Measures

Lower values are preferable for each measure shown.

NLL, Multiclass Brier and ECE were calculated over all 324 cases; Binary Brier over the 114 Boolean cases; and Score MAE over the 64 graded-score cases. Probabilities within a restricted candidate set should not be treated as calibrated probabilities of correctness in deployment.

Limitations of the Evaluation

None of the 324 reference labels has undergone human review. Agreement with these fixed references does not establish accuracy against independently confirmed answers.

The cases include related variants and templates. The row-level McNemar p-values are nominal and do not correct for dependence within clusters. They do not support generalisation to 324 independent questions. The confidence interval for the difference uses bootstrapping at the family level.

Every recorded case has source_split=train. Overlap with this model's training data has not been checked; this card consequently makes no claim that the evaluation constitutes an independent holdout.

The comparison is retrospective and draws upon results obtained at different times. It does not establish a causal training effect, broad generalisation, long-context capability or independently assessed safety. Results by question type and domain are exploratory, and improvements were not observed in every domain.

Only aggregate statistics and figures are included in this repository. Row-level CSV files and the input data are not distributed here.

Reported Evaluation Environment

The retained execution records report the following evaluation configuration. Reproduction requires an equivalent PrismL-compatible build, sufficient available VRAM and an available server port.

Download

hf download zapabobouj/Hikari-Bonsai-27B-Nimble-LoRA-GGUF --include "Hikari-LoRA-fixed-scale-PQ2_0.gguf" --local-dir .\models\Hikari-Bonsai-27B-Nimble-LoRA-GGUF

Example Server Invocation

The following Windows example reflects the recorded runtime settings. Adjust the executable and model paths to their actual locations.

.\llama-server.exe -m .\Hikari-LoRA-fixed-scale-PQ2_0.gguf -c 65536 --parallel 1 -ngl 99 -fa on -ctk q4_0 -ctv q4_0 --jinja --alias Hikari-Bonsai-27B-Nimble-LoRA-GGUF --host 127.0.0.1 --port 8080

The OpenAI-compatible API endpoint is http://127.0.0.1:8080/v1, and the model identifier is Hikari-Bonsai-27B-Nimble-LoRA-GGUF. This is a reported configuration for the single machine described above. It does not establish compatibility with every environment. Adjust context length and GPU offloading to suit the available hardware.

Licence

The card metadata retains Apache-2.0, in accordance with the licence designation already set for this Hugging Face repository. The materials examined do not establish the complete rights and permissions associated with the source model and training data. Before redistribution or commercial use, ascertain the applicable terms for both.