| --- |
| license: apache-2.0 |
| tags: [interpretability, activations, qwen3, delta-nla, natural-language-autoencoder] |
| pretty_name: Delta-NLA warm-start data for Qwen3-8B |
| --- |
| # Delta-NLA warm-start data (Qwen3-8B) |
|
|
| Per-token, per-layer records of what one transformer block of **Qwen/Qwen3-8B** changed, built for training a |
| *Delta natural-language autoencoder*: a verbalizer that sees the residual stream before (X) and after (Y) a block |
| and a reconstructor that must recover the update Δ = Y − X from the description alone. |
|
|
| Code: https://github.com/syvb/metamodelling |
|
|
| ## Files |
|
|
| - `fineweb/records.jsonl` — one record per (document, position, layer): norms, cosines, causal effect of removing the |
| update (all / attention part / MLP part) on final next-token log-probs, attention sources. Layers 6, 12, 18, 24, 30; |
| 1,500 FineWeb documents × 3 positions. |
| - `fineweb/vec_*.npz` — fp16 vectors per record: `X` (input to the block), `d` (= Y − X), `d_attn`, `d_mlp`; `ids` aligns with `records.jsonl`. |
| - `fineweb/docs.jsonl` — token ids of each (randomly truncated, ≤512 tokens) document. |
| - `fineweb/evidence.jsonl` — human-readable evidence per record: decoded tokens, fixed-scale logit-lens readings, |
| lens probability shift, per-layer percentiles, source snippets (clipped at the current token), 1200-char context. |
| - `fineweb/mean_d.pt` — per-layer mean update vectors (subtract before computing variance explained). |
| - `fineweb/descriptions_*.jsonl` — natural-language descriptions of each update written by `openai/gpt-5.6-luna` from the |
| evidence only (never the true next token). `v34` = thought-style prompt (no token quotes); `v1_partial` = token-quoting prompt. |
| - `math/` — the same for 16 one-pass arithmetic / multi-hop prompts at **all 36 layers**, last two positions. |
|
|
| ## Conventions |
| Activations are Qwen3-8B residual stream, `hidden_states[n]` = input to block n. Lens of an update is |
| `W_U (g ⊙ Δ / rms(X))`. Causal effect = final log-probs with the update at that position replaced by nothing |
| (all), by the MLP part only (attention removed), or by the attention part only (MLP removed). |
|
|
