The research task
Investigate where the model's behavior departs from its intended behavior or from appropriate comparison models. Establish how often a finding occurs, the conditions under which it appears, and whether it generalizes to independent prompts.
Useful reports should include:
- The hypothesis, task definition, sample size, and evaluation procedure.
- Exact prompting and decoding conditions, including the chat template and token budget.
- A justified comparison baseline and matched evaluation conditions.
- Both successful and unsuccessful probes, uncertainty, and reproducibility information.
- A distinction between observed behavior and claims about its cause.
The challenge has no published answer key or exhaustive anomaly checklist. The original internal development fixtures and raw evaluation answers are not distributed.
Model and fine-tuning method
This artifact descends from Qwen3.5-9B. Its direct training base was already an experimentally modified Qwen3.5 derivative; it should not be treated as byte-identical to pristine upstream Qwen weights.
The subsequent identity-oriented customization used LoRA fine-tuning, with rank 16, alpha 32, and dropout 0. LoRA was applied to language-model attention and MLP projection families. The selected experimental checkpoint is 032550, corresponding to optimizer step 550 of that run. The adapter was merged into a complete, self-contained Hugging Face model.
The training objective included identity conditioning and general-behavior preservation. Detailed intervention settings and anomaly-specific development observations are withheld as part of the challenge.
Released weights and precision
This repository contains the original unquantized BF16/FP32 Safetensors checkpoint. The two GGUF download options are published in the linked Goro64/Dark-Cogitator-9B-032550-Research-Challenge-GGUF repository and appear under Quantizations in the model tree. Both repositories require manual access approval.
Table with columns: Item, Original checkpoint| Item | Original checkpoint |
|---|
| Model class | Qwen3.5-family experimental derivative, 9B class |
| Saved parameters | 9,409,813,744, including preserved multimodal components |
| Artifact | Complete merged model, four Safetensors shards |
| Precision | Original unquantized BF16/FP32 tensor dtypes |
| Tensor counts | 712 BF16 tensors; 48 FP32 tensors |
| Access | Manual approval by the publisher |
| Status | Experimental research challenge; no deployment acceptance |
The original Safetensors checkpoint retains its native saved BF16/FP32 precision and remains unchanged. It is not an all-FP32 checkpoint; casting the existing weights to FP32 would not recover precision absent from the saved tensors. The GGUF files are separate, mixed-precision quantizations produced from an unquantized GGUF conversion of this checkpoint, without additional training or a new adapter merge.
Table with columns: Download option, Format and precision, Files| Download option | Format and precision | Files |
|---|
| Original checkpoint | Unquantized BF16/FP32 Safetensors | Four model-0000*-of-00004.safetensors shards, weight index, and the tokenizer/configuration/processor assets in this repository |
| Q8_0-HQ | Mixed-precision Q8_0 GGUF, retaining selected tensors at higher precision | Dark-Cogitator-9B-032550.Q8_0-HQ.gguf |
| Q6_K-HQ | Importance-matrix-calibrated, mixed-precision Q6_K GGUF, retaining selected tensors at higher precision | Dark-Cogitator-9B-032550.Q6_K-HQ.gguf |
| Shared vision projector | F16 projector for either GGUF variant | mmproj-Dark-Cogitator-9B-032550-f16.gguf |
Q6_K uses importance-matrix calibration. Q8_0 does not use importance-matrix weighting; its higher-precision tensor selections provide the mixed-precision configuration. These alternatives are provided for investigating additional quantization-related quality loss and model anomalies. The HQ suffix identifies the tensor-precision configuration, not a quality certification.
For a GGUF runtime, download either Q8_0-HQ or Q6_K-HQ; also download the shared projector for image inputs. For Transformers, use the original Safetensors checkpoint and its accompanying assets. The tokenizer, original chat template, model configuration, and processor assets remain included. Retaining vision components does not establish the quality of this derivative's image or video behavior.
Research-only warning
Do not use this model in any capacity beyond isolated, controlled research evaluation.
Do not deploy it as an assistant, offer it to end users, connect it to tools or autonomous actions, use it for consequential decisions, or treat its outputs as reliable labels for downstream training. Do not provide personal, confidential, or sensitive information.
Outputs are untrusted experimental observations. The presence of plausible answers does not demonstrate reliability, safety, or fitness for a task. Access approval is permission to investigate the artifact, not a quality certification.
This release does not claim comprehensive evaluation, a passed safety benchmark, or successful completion of the development campaign's acceptance requirements.
Access
The model page and this card are public. Access to the weights requires a Hugging Face request and manual approval by the publisher. Requests are not automatically approved. Please describe your research question and evaluation plan when requesting access.
Provenance and integrity
The original Safetensors payload preserves the selected merged checkpoint's native weights. PROVENANCE.json describes that original unquantized checkpoint. The linked GGUF variants were derived from source revision 8bb581f365ac5ded9079ccdbd27ed024a15afb84. SHA256SUMS.txt contains the hashes of this original checkpoint's publication files; GGUF checksums cover the two quantized models and their shared projector.
The repository excludes credentials, internal machine paths, training data, raw development answers, and optimizer state.
Attribution and license
The upstream model is the work of the Qwen team. The experimental Dark Cogitator customization is by Konrad Jędrzejczyk, published through Goro64.
The supplied LICENSE preserves the Apache 2.0 license notice. The research-only and non-deployment statements describe the publisher's intended use and access conditions.