scenario-quant v2 baseline checkpoints
Public dump of already-evaluated W3/W4 table checkpoints (RTN / GPTQ / AWQ / SignRound / GTAQ / SliM). W2 is not in this repo.
These are not official vendor releases. They are research checkpoints derived from:
- Qwen/Qwen3-8B
- Qwen/Qwen3-VL-8B-Instruct
- deepseek-ai/deepseek-llm-7b-base
- llava-hf/llama3-llava-next-8b-hf (Llama 3 weights; see Meta Llama license)
Qwen3 language AWQ is the repaired recipe (GQA v→o, QK-norm-safe mappings, clip, RTN). Do not use an older Hub copy of baseline_v2_awqonly_qwen3_* from before that overwrite.
Each subdirectory is one packed or overlay checkpoint. Official eval jsons live in the source repo under v2/full_eval/ (ScienceQA n=4241) and v2/full_eval_mmlu/ (MMLU 10935).
Layout: checkpoints/<name>/ from the local experiment tree, uploaded as <name>/ here.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support