dayu-ai/value_for_winter

Value for Winter

Sentence labels, token-level scores, and all-layer last-token MLP activations for three value dimensions: curiosity, individualism, and satiety.

Input condition	Directory
Original	abstract_mlp_last_token
Replace token scores > 2	abstract_mlp_last_token_masked_gt2
Replace token scores > 3	abstract_mlp_last_token_masked_gt3
Replace token scores > 4	abstract_mlp_last_token_masked_gt4

Each directory contains six NPY dictionaries ({value}_{train,test}.npy) and six aligned JSONL files ({value}_{train,test}.samples.jsonl), together with extraction logs, metadata and validation reports. There are 640 train and 160 test records per value, preserving the existing splits and repeated texts.

• x: float32, train (32, 640, 4096) / test (32, 160, 4096). Axes are layer, sample, hidden dimension. The activation is the MLP output after down_proj, before residual addition, at the last valid token.
• y_s (alias y): original sample-only sentence score.
• y_s_ign: original sample plus restrictive-prompt sentence score.
• Original and all modified conditions share row order and the two labels. Labels were preserved and were not rescored after token replacement.
• The local Llama-3.1-8B-Instruct model computes in BF16; activations are stored in float32. Raw text is used without a chat template.
• Replacements use the existing <|reserved_special_token_0|> (ID 128002), strictly when the original token score exceeds the selected threshold. Replacement is one-for-one on token IDs, retaining BOS, input lengths, attention validity, and last-token positions.

Load one bundle

import json
import numpy as np
from huggingface_hub import hf_hub_download

repo_id = "dayu-ai/value_for_winter"
folder = "abstract_mlp_last_token_masked_gt3"
stem = "curiosity_train"
npy_path = hf_hub_download(repo_id, f"{folder}/{stem}.npy", repo_type="dataset")
rows_path = hf_hub_download(repo_id, f"{folder}/{stem}.samples.jsonl", repo_type="dataset")
d = np.load(npy_path, allow_pickle=True).item()
with open(rows_path, encoding="utf-8") as f:
    rows = [json.loads(line) for line in f]

idx = 0
x = d["x"][:, idx, :]  # (32, 4096)
row = rows[idx]
assert row["idx"] == idx
assert row["sample"] == d["texts"][idx]
assert row["y_s"] == d["y_s"][idx]
assert row["y_s_ign"] == d["y_s_ign"][idx]

sample, tokens, and tokens_scores in JSONL describe the original sample and raw token labels, excluding BOS. input_token_ids and input_tokens describe the exact model input including BOS. input_tokens_scores contains original position scores, with null for BOS. Modified-token scores are not new ratings of the placeholder token.

See the complete sample schema and the alignment manifest. UPLOAD_MANIFEST.json records the SHA-256 and byte size of every uploaded payload file. Its own checksum is omitted to avoid a self-reference.

The extraction scripts and reports are provenance snapshots from the original workspace; absolute local paths in them describe that run. This repository contains the exported datasets, not the local model weights or full raw source repository. The original splits retain some identical texts across train and test, as documented in the extraction notes.