araujota/activation-rag-telemetry-backbone-v1

Activation RAG Telemetry Backbone v1

This dependency bundle contains the exact custom checkpoints used to collect telemetry for araujota/activation-support-fusion-v1:

• qwen3_rmt_joint_memory_latest.pt: Qwen3-4B RMT joint-memory adapter/checkpoint;
• topk_sae_latest.pt: layer-7 TopK sparse autoencoder;
• feature_manifest.json: complete SAE feature catalog.

It does not include Qwen3-4B. Obtain the base model from Qwen/Qwen3-4B and comply with its license and terms.

Checksums

5399ee81a8c7ec66d33c32fbbb4553ab8e43049851a7c2a8ce7b8e6cf0b23543  qwen3_rmt_joint_memory_latest.pt
3f24623680aaa3bcd85ffadc8988290f62d4e00d3af08b69d7bc6fba1d7e45a4  topk_sae_latest.pt
8c9d9699a7c3a97aa543839d1ba2bae2f6c2d24f9a1fe8b60d56d883ed9cc988  feature_manifest.json

Download

hf download araujota/activation-rag-telemetry-backbone-v1 \
  --local-dir activation-rag-telemetry-backbone-v1

Then pass the paths explicitly to the capture runner:

python scripts/capture_qwen_l24_resid_pre_sharded.py \
  --requests work/l24-requests/support-residual-capture-requests.jsonl \
  --out-dir work/l24-capture \
  --model-path /models/Qwen3-4B \
  --joint-checkpoint activation-rag-telemetry-backbone-v1/qwen3_rmt_joint_memory_latest.pt \
  --sae-checkpoint activation-rag-telemetry-backbone-v1/topk_sae_latest.pt \
  --device cuda

Layer-7 Core245 collection uses the same checkpoints and the exact feature IDs in the fusion artifact's feature-manifest.json. See the activation-rag README for prompt, position, pooling, residual, and scorer contracts.

Security And Compatibility

The two .pt files are legacy PyTorch checkpoint containers. PyTorch pickle files can execute code when loaded; download only from this repository, verify the SHA-256 hashes, and use torch.load(..., weights_only=True) where the consumer supports it. The reference RMT runner may require its original trusted checkpoint loader.

Telemetry compatibility is exact rather than approximate. Changing the base model, adapter, SAE, tokenizer, layer, capture site, prompt, token window, decision position, or pooling invalidates the published reranker calibration.

Intended Use

This bundle is for research reproduction and development of prefill-only activation telemetry. It is not a standalone language model, reranker, medical model, or legal model. No benchmark text, qrels, raw hidden states, or telemetry caches are included.