srt-nla-targets-gemma2-2b-v1 — 30K (activation, text) pairs from Gemma-2-2B L19

The training and evaluation corpus for srt-nla-av-gemma2-2b-v1, the third-backbone replication of RiverRider/srt-nla-av-v1 on Gemma-2-2B. Each example is a (hidden activation, source text) pair where the activation is the last valid token's L19 hidden state of a Gemma-2-2B continuation of length 64.

Card metadata

Backbone	google/gemma-2-2b
Layer	19 (73% depth, mirrors Qwen's L20/28 and Llama's L20/28)
Token position	Last valid token (attention-mask determined)
Sequence length	64 tokens
Dtype	fp32 (activation), str (text)
Activation dim	2304
N targets	30,000 (seed=1)
Pool size (paper anchors)	2,000
Anisotropy ‖μ‖	≈ 156.3 (~3× Qwen-2.5-7B, ~22× Llama-3.2-3B)
SHA-256 (full file)	129f1322…fba9d

Files

File	Size	Notes
targets_L19_seq64_30k_seed1.pt	~17 GB	Full (sequences, activations, meta). Use with weights_only=False.

Schema

obj = torch.load(path, weights_only=False)
# obj["sequences"]:   list[Tensor]    — generated token id tensors
# obj["activations"]: list[Tensor]    — fp32, L19 hidden states (per sample, shape (T_i, 2304))
# obj["meta"]:        {"backbone_id": "google/gemma-2-2b", "extraction_layer": 19, "d": 2304, "seed": 1, "temperature": 1.0, "top_p": 1.0}

Note: unlike the Qwen and Llama target packs which store (N, T) / (N, T, d) dense tensors, the Gemma pack stores per-sample lists. This follows the post-902b746 sampler's preferred format and avoids padding-mask ambiguity at evaluation time.

Reproduction

python scripts/sample_targets.py \
  --backbone google/gemma-2-2b \
  --layer 19 --seq-len 64 \
  --num-sequences 30000 --batch-size 16 \
  --dtype bfloat16 --seed 1 \
  --out artifacts/nla/gemma2_2B/targets_L19_seq64_30k_seed1.pt

The activation extractor (scripts/sample_targets.py) is backbone-agnostic; no code changes were needed to switch from Llama to Gemma. Gemma-2-2B has distinct bos_token_id=2 and eos_token_id=1 — explicitly not the Qwen-style trap that produced the constant-target bug fixed at commit 902b746. The companion gold-pair builder is similarly parametric: scripts/build_gold_pairs.py --backbone google/gemma-2-2b --max-len 64, which yields 29,952 pairs (48 token-budget skips at this length).

Anchors derived from this dataset

Computed on a 200-target held-out slice (see oracle_ceiling_M200.json released alongside the model card), pool=2000:

anchor	raw fve_nrm	centred
replay (re-encode)	0.799	0.713
NN-retrieval (pool=2000)	0.781	0.653
paraphrase best-of-8 (Gemma)	0.720	0.598 ← headline ceiling
random floor (off-diagonal)	0.675	0.498 ← floor

Centring subtracts the per-coordinate pool mean before cosine; on Gemma-2-2B ‖μ‖ ≈ 156. Note that raw greedy verbalizer fve (0.664) sits below the raw random floor (0.675) — the strongest empirical case across the three backbones for the non-optionality of centring on high-anisotropy backbones.

Intended use

• Training and evaluating activation verbalizers / decoders for Gemma-2-2B L19.
• Building hard-negative pools for InfoNCE-style activation→text losses.
• Cross-backbone studies of mid-layer representation geometry (Qwen vs Llama vs Gemma — three labs, three architectures, three anisotropy regimes).

Out-of-scope

• Numbers do not transfer to other backbones, sizes, or layers without recomputing μ, the floor, and the ceiling.
• Texts are model-generated continuations of seed prompts, not human-written; do not treat as a natural-language corpus for unrelated NLP work.

Citation

See srt-nla-av-gemma2-2b-v1 model card.