Nanonla Qwen3 8B L24 Data Full: Neural Network Activation Vectors

syvb published the 'Qwen3-8B NLA — FULL parquets' dataset on Hugging Face on 2026-06-19. It contains neural network activation vectors recomputed for the Qwen3-8B model. The data is split into three configurations: av_sft, ar_sft, and rl, each with a different prompt schema.

Use Cases

• Analyzing layer-24 residual activations based on the 'activation_vector' column mentioned in the description.
• Comparing model behavior across different training stages based on the described 'av_sft', 'ar_sft', and 'rl' configurations.
• Investigating the impact of prompt schemas on activation patterns, as the description notes separate configs for this reason.

Strengths

• Data is specifically generated for the Qwen3-8B model, a known large language model.
• The 'activation_vector' column was regenerated, suggesting a focus on data consistency.
• Provides three distinct configurations (av_sft, ar_sft, rl) for comparative analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

syvb on Hugging Face.

Collection Method

Activation vectors were regenerated from the Qwen3-8B model.

Freshness

Last updated 2026-06-19 19:08:07; freshness should be verified.