syvb/qwen3-last-position-patching

Qwen3 last-position residual patching

This is the data from an experiment on how "self-correcting" the Qwen3 base models are (0.6B, 1.7B, 4B, 8B, 14B). For a held-out FineWeb document cut at position T, the residual stream at the last position is replaced at one site with an ablated value. We then measure how far every later site's residual stream at that position moves from the clean run, and how much the next-token prediction changes.

Code: https://github.com/syvb/act-experiments (branch self-correction). The README there defines the metrics, and RESULTS.md has the findings.

• Sites: 0 = token embeddings, i = residual stream after i blocks (before the final norm).
• Ablations:
  • mean: per-token mean residual over other contexts ending in the same token.
  • resample: the residual of one other real context (a different doc) ending in the same token.
  • global: the mean over all tokens.
• Data: HuggingFaceFW/fineweb sample/10BT/000_00000.parquet. Docs are truncated to 512 tokens and positions [64, 512) are used. Means come from the first ~34k docs of the file (10M positions). Test samples come from the last 10k docs, iterated from the end of the file; doc indexes that reversed order.
• Precision: means were computed in bf16, and patching in fp32 (no TF32). Checks on 0.6B and 14B found the bf16-vs-fp32 gap in the means is 70–2000× below their sampling noise.

Files (one set per model size)

means_{size}.pt (torch): token_ids [K], counts [K] (occurrences; only tokens with ≥ 500 are kept), means [n_sites, K, d] fp32, global_mean [n_sites, d], model, dtype.

metrics_{size}.npz (numpy): per-sample results for 3,000 samples, with ablation order ablations.

• doc, T, token, donor_doc, donor_T
• spread [n, site] = |clean − token mean|²; donor_gap [n, site] = |clean − donor|²
• sq [n, abl, patch site, measured site] = |patched − clean|²
• proj [...] = ⟨patched − clean, injected⟩
• pull [...] = ⟨patched − donor, clean − donor⟩
• attn_repair, mlp_repair [n, abl, patch site, block] = −⟨Δ block output, injected⟩
• kl [n, abl, patch site] = KL(clean ‖ patched) of the next-token distribution
• dloss [n, abl, patch site] = increase in loss on the true next token

acts_{size}.pt (torch, bf16 activations):

• ids: list of token-id tensors, each holding the context, the final token, and the true next token
• clean [n, n_sites, d]: clean last-position residuals of every sample
• patched [32, abl, 1 + n_sites, n_sites, d]: every patched run of the first 32 samples (row 0 = clean, row 1 + s = patched at site s)
• also doc, T, model, ablations

metrics_robust_{size}.npz: a separate 1,000-sample pass. It has the same fields, plus sq_no_massive / spread_no_massive, which leave out the 5 dims with the largest |mean| (massive_dims). In Qwen3-4B and 8B, the patch switches a massive activation on or off in a few samples: |patched − clean| reaches 4,000–13,000 in those dims from site 17 on, where clean values are normally below ~200. In the main run, for patches at sites 1–16, 3 of the 3,000 samples carry 74% (4B) and 93% (8B) of the mean-ablation perturbation measured at sites 17–25. They also distort averages of proj (8B's retained share at 75% depth is 0.83 with them and 1.31 without). Use the _no_massive fields, or drop those samples, when averaging. In every model except 0.6B, the massive dims also inflate final-site disruption for late patches by up to ~25%.

metrics_noise_14B.npz: the noise run on 14B (run_noise.sh). It has the same fields as metrics_robust_{size}.npz, for the same 3,000 samples and donors as metrics_14B.npz. The conditions are mean, resample, add_context_x{s} and add_random_x{s} for s = 0.25, 1, 4 and 16. add_context adds s × (donor − token mean) to the clean residual, and add_random adds a random direction of the same size. Plots and a summary are in plots_noise/, and the log is in logs/noise.log.

steer/: a 0.6B steering prototype, run on CPU.

• vectors_0.6B.pt: topics (weddings, football, cooking) and vectors [topic, n_sites, d]. Each vector is the mean of (residual − token mean) after the first mention, in 200 FineWeb docs per topic from the start of the file.
• steer_0.6B.npz: one row per (sample, topic), with on_topic, topic, doc, T, token, clean_topic_logp and spread. Conditions are steer_x{s} and random_x{s} for s = 1 to 32. Per condition, with the same shapes as above: sq, proj, concept (= ⟨patched − clean, topic vector at the measured site⟩), kl, and dtopic (the change in log P of the topic words). vector_dot [topic, site, site] holds the dot products of the topic vectors.
• Plots and a summary are in plots_steer/.

check_means_14B_{bfloat16,float32}.pt: 14B means over 300k positions in both precisions, used for plots/check_means_14B.txt.

plots/ (main run) and plots_robust/ (robust pass) hold heatmaps per model, the cross-size figure, summary.txt, and key_numbers.txt. logs/ holds the run logs.

Licensing: the token ids are derived from FineWeb (ODC-By 1.0), and the Qwen3 models are Apache-2.0.

Loading

import numpy as np, torch
from huggingface_hub import hf_hub_download
f = np.load(hf_hub_download("syvb/qwen3-last-position-patching", "metrics_4B.npz", repo_type="dataset"))
acts = torch.load(hf_hub_download("syvb/qwen3-last-position-patching", "acts_4B.pt", repo_type="dataset"))