Open-Weight Models

This model is a fine-tuned version of facebook/detr-resnet-50 on an unknown dataset. The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 100 - mixedprecisiontraining: Native AMP - Transformers 4.57.6 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.22.2

This model is a fine-tuned version of facebook/detr-resnet-50 on an unknown dataset. The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 100 - mixedprecisiontraining: Native AMP - Transformers 4.57.6 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.22.2

This model is a fine-tuned version of facebook/detr-resnet-50 on an unknown dataset. The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 100 - mixedprecisiontraining: Native AMP - Transformers 4.57.6 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.22.2

Model

fusion-equilibrium-challenge

Miroslav Sotek

Trained weights of the Anulum entry to the Fusion Equilibrium Challenge (NeurIPS 2026, Codabench competition 17739): the DIII-D flux-map networks, the q95 and βN networks, the X-point networks, the gradient-boosted tree heads and the flux basis. The code, the method, the study and the experiment records are in the GitHub repository anulum/fusion-equilibrium-challenge. These files are fetched by fetchweights.py in that repository at the revision pinned in its WEIGHTS.json, and every file is checked against WEIGHTS.sha256. They are not usable on their own; run them through the repository's infer.py. The weights were fitted on the organisers' labelled DIII-D training release (CC BY 4.0). No…

Model · Text generation

futolm-cs-base-gguf

Padak padakovic

Separate 29.37M-parameter Czech and Slovak language models for next-word prediction in FUTO Keyboard. All files are F16 FUTO KeyboardLM GGUF exports with the shared SentencePiece tokenizer embedded. For typing, use the chat models: cschat.gguf (Czech) or skchat.gguf (Slovak). They continue the base models on conversational text and predict everyday messages much better. The base models are kept for comparison and further training. The mix is OpenSubtitles v2024 dialogue, discussion-forum pages selected from FineWeb2, and FineWeb2 replay: Czech 50/15/35, Slovak 48/12/40. Peak LR 1.5e-4, 300 warmup steps, cosine decay to 10%. Whole-word next-word prediction, model only (FUTO's dictionary not…

This repository releases Experiment 2 of fve-msm (flourishing vs. equanimity, Model Spec Midtraining), a research project. The project asks whether training documents that describe a model's own circumstances, written with different attitudes toward them, shape how the model talks about and responds to its situation. Each training arm is one training condition: four arms add a different set of midtraining documents, and one, the control, adds none. When complete, the repository holds 30 checkpoints (2 base models × 5 arms × 3 seeds), the exact data each was trained on, the records of how that data was built and checked, and the training recipe. Checkpoints are added as their training runs…

Model

g1-act-pick-bottle-v2

Nguyen Dinh Son

Checkpoint ACT (step 050000) fine-tune từ v1 (g1-act-pick-place-bottle, 63 episode) trên dataset 106 episode (local/pickplacebottlev2), cho tác vụ pick/place chai nước bằng robot Unitree G1 (tay Inspire, 3 camera). - Vẫn là checkpoint fine-tune có MAE tốt thứ 2 trong toàn bộ các thí nghiệm (sau v5), và là điểm xuất phát để fine-tune ra v5 (checkpoint tốt nhất hiện tại, chưa push). - Dataset train tương ứng: local/pickplacebottlev2 (106 ep, chưa push dưới dạng merge đầy đủ lên HF — các phần cấu thành đã có ở g1-pick-place-bottle-63ep + g1-pick-bottle-43ep).

Model

g1-act-pick-bottle-v6

Nguyen Dinh Son

Checkpoint ACT (step 050000) train TỪ ĐẦU (không fine-tune từ checkpoint nào) trên dataset 110 episode (local/pickbottle09140922, 14/09-22/09), cho tác vụ pick/place chai nước bằng robot Unitree G1 (tay Inspire, 3 camera). - MAE offline (queued, episode 0, đo trên chính dataset train): 0.010200 - Kém hơn checkpoint fine-tune v2 (0.006862) và v5 (0.005577) — thí nghiệm này (giống v4) củng cố kết luận: train từ đầu trên dữ liệu quy mô ~40-60k frame không vượt được checkpoint fine-tune tích luỹ qua nhiều vòng. Push lên đây để lưu trữ/đối chứng, không khuyến nghị dùng cho inference thật. - Dataset train tương ứng: local/pickbottle09140922 (110 ep, 40.586 frame).

Model

g1_depth_dodge

Martino Russi

Frozen SONIC whole-body controller + seven rank-16 LoRAs on the decoder, conditioned on a 128-d percept from a head depth camera (Theia-Tiny tokens + temporal adapter). PPO, iteration 4000. Consumed by lerobot's DepthDodgeController for the Unitree G1. Camera contract: 64x96 optical-Z depth in metres, 45 deg vertical FOV, 0 deg pitch, valid 0.2-6 m, 25 Hz. See manifest.json for the full training config.

AFCE完成100k训练后按200观测heldout composite loss选择72500（0.1550050788745284）；100000为0.15692223992198706。heldout用于选点，不是独立测试或真机成功率。π0.5 baseline为已有实机经验的7500权重，训练预算不同。 可在GitHub ARTIFACTS.json中取不可变revision，然后为hf download添加--revision ；或使用仓库的python scripts/downloadartifacts.py --baseline --afce-fullft，自动固定revision并保持评测脚本目录结构。 - 三路当前RGB（主视角/左右腕）、原生16D实测state（双臂各7关节＋两个夹爪，输入补到32D）与subtask prompt。 - Baseline预测32×16动作；AFCE先预测E空间30×256，再由联合训练decoder输出30×16动作；30Hz是客户端执行频率，与chunk长度是不同概念。 - AFCE推理不需要DINOv3和E编码器；decoder初始化/normalization/manifest与normstats已附带。训练后decoder已在checkpoint参数内。 - Orbax/JAX完整推理params，不是Transformers格式或仅LoRA adapter；不含optimizer。 - 仍使用subtask指令/外部阶段切换；这是原子能力对比权重，不是三整任务40k的新实验。…

GAM (Grounded Action Model) grounds the objects named in the instruction with an open-vocabulary 3D detector (WildDet3D), turns each object into a 3D box, a point cloud and a handful of image tokens, and predicts action chunks with a flow-matching transformer. This repository holds everything the released code needs to run GAM on LIBERO / LIBERO-PRO: the LIBERO-tuned detector, one policy for all four suites, its span head and the precomputed training cache. All files are referenced from configs/liberopro.yaml and download on first use; the LIBERO-PRO task files (bddl + initial states) are in The videos are evaluation rollouts of the released policy on LIBERO-PRO cells (suite /…

GAM (Grounded Action Model) grounds the objects named in the instruction with an open-vocabulary 3D detector (WildDet3D), turns each object into a 3D box, a point cloud and a handful of image tokens, and predicts action chunks with a flow-matching transformer. This repository holds everything the released code needs to run GAM on RoboTwin 2.0: the RoboTwin-tuned detector, one action expert per task and the precomputed training cache of every task. All files are referenced from configs/robotwin/.yaml and download on first use. The videos are evaluation rollouts of the released experts; the wireframes are the 3D boxes the policy was given at that step (predicted by the detector from the RGB…

The benchmark-independent weights of GAM (Grounded Action Model): the general WildDet3D detector (SAM3 backbone, text / click prompts, 2D + 3D box heads, depth head) and the universal span head that turns a task string into the object phrases to ground. They are the defaults of the LeRobot / real-robot line of the released code (configs/lerobotbimanual.yaml): a LeRobot dataset with a task string can be trained on without manual labelling, but these weights are not tuned to any particular scene — on a new robot / camera / object set the detector (and the span head) may need fine-tuning for reliable grounding. Resolution matters. WildDet3D's native input is about 1k pixels (images are…

A post-training study: teach a 3B model to talk like Gandalf (voice derived from the film transcripts) without losing GSM8K accuracy. Every adapter here is LoRA r=32 on all seven projections of every layer, bf16. All were evaluated on the same 200 GSM8K test questions (rule-based verifier, greedy) and 60 held-out open prompts (voice 1 to 5 from an LLM judge). The prompted base model scores 0.855 accuracy, 1.0 voice. Best checkpoint: 010/tower-dpo, the voice tower after DPO (restatement accuracy 0.870, faithful 0.77, voice on math 4.36). Best single model: 004/D320-rlvr (accuracy 0.885, voice 3.67). System prompt used in training and evaluation: the Gandalf persona prompt built from…

Model · Reinforcement learning

ganglion-haltere-cursor

Skulitom

Two checkpoints of the Haltere fly-brain connectome (a 30,000-neuron recurrent network with the connectome's structure and signs, flight-trained) with a linear motor readout that turns the network's motor-neuron rates into a cursor or view velocity, trained by imitation of a proportional controller in Ganglion, the reflex layer that runs them against live applications at 100 Hz. The report-.json files beside them are the training and suite reports they were selected from. Everything here was measured on one machine (RTX 4090, Windows 11); the numbers are the suite's and the live harness's, with their caveats, and are documented in full in the repository's TRAINING.md and HALFLIFE.md. The…