Instructions to use wenyii/cdmt-checkpoints-extra with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use wenyii/cdmt-checkpoints-extra with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("wenyii/cdmt-checkpoints-extra", device_map="auto") - Notebooks
- Google Colab
- Kaggle
CDMT additional checkpoints
Companion to wenyii/cdmt-qwen2.5-3b. Terminal checkpoints for the scale, model-family, and teacher-choice
experiments, in Hugging Face format (one subfolder per run, seeds 11/22/33):
| Subfolders | Base model | Training |
|---|---|---|
qwen2.5-1.5b-search-{sft,grpo}-s* |
Qwen/Qwen2.5-1.5B-Instruct | Search, 10k prompts (SFT 1,250 steps / GRPO 250 steps, lr 1e-6) |
qwen2.5-7b-search-{sft,grpo}-s* |
Qwen/Qwen2.5-7B-Instruct | same |
gemma3-4b-search-sft-s* |
HPC adapted Gemma text-only base (chat template in gemma-3-4b-it-cdmt-template/) |
1,000 Qwen2.5-72B teacher Search trajectories, lr 1e-5, 1 epoch |
llama3.1-8b-search-sft-s* |
meta-llama/Llama-3.1-8B-Instruct | same as Gemma |
qwen2.5-3b-search-sft-xteacher-{qwen72b,llama31}-s* |
Qwen/Qwen2.5-3B-Instruct | Search SFT on trajectories from the named teacher (matched prompts) |
qwen2.5-3b-search-sft-legacy-full-s* |
Qwen/Qwen2.5-3B-Instruct | earlier Search-SFT recipe (15 steps) |
gemma-3-4b-it-cdmt-template/ holds only config, tokenizer, and chat-template files (no weights); combine with
the adapted Gemma text-only checkpoint used on the original HPC. A direct download of google/gemma-3-4b-it has different shard hashes; the template is not a conversion or a replacement for the historical Base weights. base_model in the checkpoint catalogue records model lineage, not the verified bytes of that Base. Licenses follow the base model of each subfolder: Qwen2.5-1.5B/7B-Instruct Apache-2.0; Qwen2.5-3B-Instruct Qwen Research License (non-commercial); Gemma derivatives the Gemma Terms of Use; Llama derivatives the Llama 3.1 Community License (Built with Llama). SFT data includes Qwen2.5-72B-Instruct outputs (Built with Qwen).
Supplemental checkpoint groups
| Subfolders | Training |
|---|---|
qwen2.5-{1.5b,3b,7b}-codedebug-{sft,grpo}-s* |
CodeDebug scale panel, seeds 11/22/33 |
gemma3-4b-search-grpo-s* |
Gemma Search GRPO, seeds 11/22/33 |
qwen2.5-3b-search-grpo-legacy-full-s* |
Legacy Search GRPO used in the earlier nonce panel |
Gemma is provided under and subject to the Gemma Terms of Use found at https://ai.google.dev/gemma/terms. Qwen checkpoint license notices follow their respective upstream model repositories.
Earlier external Search-panel checkpoints
These 18 Qwen2.5-3B checkpoints (Code/Math/Search, SFT/OPD, seeds 11/22/33) are the exact weights used for the earlier Bamboogle, 2Wiki and MuSiQue evaluations. Their folder names contain legacy-external. Use checkpoints.json to map evaluation model tags to subfolders. The dataset repository contains the corresponding eval_results/external_panel trajectories. The weights follow the Qwen2.5-3B-Instruct Research License.