Oratis's picture
补齐39份原始预处理配置并完成25份研究模型归档核验
65a20d4 verified
|
Raw History Blame Contribute Delete
5 kB
---
license: cc-by-nc-4.0
base_model:
- Qwen/Qwen3-VL-4B-Instruct
- facebook/dinov2-small
base_model_relation: adapter
tags:
- research-archive
- lora
- inverse-dynamics
- gameplay
- safetensors
---
# Diogenes research models archive
Historical project-closeout archive of **13 Qwen3-VL LoRA variants and 12 trained
inverse-dynamics heads**. These are experimental research artifacts, with no new
training, benchmark run, certification, product release or inference service.
All 25 model payloads have been uploaded and verified by exact SHA-256 and size,
with anonymous pinned downloads. `MANIFEST.jsonl` records the complete file set.
## What is preserved
- `adapters/<variant>/adapter_model.safetensors` and `adapter_config.json`: original
LoRA tensors and configs for **Qwen/Qwen3-VL-4B-Instruct**. Rank 32; targets and
hyperparameters are in each config. Six original compressed bundles were unpacked
without changing tensor bytes. Seven directory variants preserve their files.
Each variant also preserves its original chat template, tokenizer settings and
image processor settings. Load the upstream vocabulary separately.
- `idm/<variant>/head.safetensors`: all **58 trained head tensors** from each original
checkpoint, with the original `head.` key prefix. Each tensor was compared byte
for byte against its checkpoint storage and round-tripped through the Safetensors
parser. `config.json` retains model architecture, data settings, training recipe,
gates and the actual saved training step.
- `idm/BASE_TENSOR_HASHES.json`: names, shapes and SHA-256 for the **175 frozen DINOv2
ViT-S/14 trunk tensors**. They match across all twelve original checkpoints.
Frozen downloaded trunk weights are omitted; use the upstream encoder and compare
these digests when reproducing an experiment.
- `adapters/MANIFEST.json` and `idm/EXPORT_MANIFEST.json`: source and export integrity
records. A head-only Safetensors file is a documented transformation, not a claim
that the original full checkpoint or its file hash was preserved byte for byte.
Base models, downloaded benchmark frames, upstream tokenizer vocabularies, runtime environments,
private operational logs and account credentials are not included.
## Provenance and limits
The LoRA variants belong to the game-observation method-validation experiments on
[Markov AI Gaming500](https://huggingface.co/datasets/markov-ai/gaming-500-hours).
Most IDM heads use Gaming500 conversions; the three `idm-d2e-*` variants use the
[Open World Agents D2E-480p](https://huggingface.co/datasets/open-world-agents/D2E-480p)
conversion with Rainbow Six held out. Both upstream dataset cards currently declare
**CC BY-NC 4.0**. These are third-party sources, not Diogenes self-captured data.
The historical source and base-model revisions were not pinned: current revisions
in `UPSTREAM_REFERENCES.json` are reference information only.
Original experiment metadata retains `is_trainable=false` / method-validation
routing, and explicit method-validation overrides were used for these experiments.
Publication does not change those flags or certify product-training eligibility.
Some LoRA recipes used Qwen teacher labels and hybrid ground truth. The separate
Gemini evaluation-only labels are not part of this model-weight package. No uniform
training recipe, successful quality gate, generalization result or commercial-use
permission is claimed. Some checkpoints stop before their configured target step.
The IDM head consumes the frozen encoder's features over a non-causal five-frame
window. The config records resolution, stride and action scaling; do not assume
all source datasets have equivalent pointer labels. Gaming500 conversions use
absolute-pointer differences, while D2E conversion uses raw HID deltas and an
explicit frame-time grid. Preserve that distinction when comparing results.
## Use
Select a variant directory explicitly. For a LoRA, load the recorded Qwen base and
attach that directory using PEFT. For an IDM head, construct the model from its
config and the published [Diogenes source](https://github.com/CuddlerAI/Diogenes-Archive),
load the upstream DINOv2 encoder, then load the head tensors with the recorded prefix.
This aggregate repository is not a standalone Transformers model and does not
provide an automatic inference endpoint. Closeout validation checks preservation
and serialization, not a fresh inference run or numerical quality evaluation.
## Terms and attribution
Diogenes/CuddlerAI publishes its own contributions in these research-model artifacts
under **CC BY-NC 4.0**. This noncommercial research distribution does not relicense
upstream Qwen/DINOv2 code or grant rights in third-party game imagery, characters,
trademarks or source datasets. Qwen and DINOv2 are separately Apache-2.0; see `NOTICE`
and the included upstream license text. Project source code is separately published
under Apache-2.0 in GitHub. Keep source attribution and documented transformations.