tchaitanya0109/anlp-hard-test-generations

Hard-test neutral rewrites from eight LED models

Neutral rewrites of 1,187 strongly sided validation/test articles, produced by eight allenai/led-base-16384 neutralisers from the ANLP team project. The originals and the Nemotron n2 targets are included as reference rows.

The articles come from Raunak's hard-test (branch data @ 69ebfdf of github.com/RaunakSeksaria/ANLP-Project): validation/test articles that the POLITICS-v2 classifier places most confidently on their own side (top 20 % by own-side logit margin), with no training articles. All 1,187 articles here fit within the models' 512-token input (title </s> body), so every model read each article in full: 711 from the dedup release and 476 from full1024.

Model key	Model	HF repo / subfolder
baseline_lambda0	generation-only baseline (λ 0, masked mean)	RaunakSeksaria/anlp-neutral-led / side-neutral-full/lambda0
ideology_standard_m04	first token, standard triplet, margin 0.4, ideology negatives	tchaitanya0109/anlp-neutral-led-ideology-first-token / standard-margin04-lambda035
ideology_attraction_m02	first token + attraction, margin 0.2, ideology negatives	same repo / attraction-margin02-lambda035
ideology_attraction_m05	first token + attraction, margin 0.5, ideology negatives	same repo / attraction-margin05-lambda035
publisher_standard_m02	first token, standard triplet, margin 0.2, publisher negatives	tchaitanya0109/anlp-neutral-led-publisher-first-token / standard-lambda035
publisher_attraction_m02	first token + attraction, margin 0.2, publisher negatives	same repo / attraction-lambda035
ideology_attraction_m03_b05_t03	first token + attraction (β 0.5, τ 0.3), margin 0.3, ideology negatives	tchaitanya0109/anlp-neutral-led-ideology-first-token / attraction-margin03-beta05-tau03-lambda035
ideology_attraction_m03_b06_t05	first token + attraction (β 0.6, τ 0.5), margin 0.3, ideology negatives	same repo / attraction-margin03-beta06-tau05-lambda035

The attraction models use β 0.3 and τ 0.5 unless stated. The last two were trained on Raunak's side_neutral_full training split (12,484 triplets, the same split as the baseline); the other ideology models on Manas's side_neutral split.

Exact weights used (SHA-256 of each model.safetensors), EOS rates and mean output lengths are in summary.json.

How the rewrites were made

Exactly the training-time validation decoding of the project code (branch manas): input title </s> body (title and body .strip()-ed, as in prepare_side_neutral.py), maximum input length 512 tokens, global attention on token 0, attention window 512, load_led_triplet + generate_with_cache, greedy (num_beams=1, do_sample=False), max_new_tokens=512, fp32, one article at a time, on RTX 2080 Ti GPUs (IIIT Ada). Checks: the 711 dedup inputs equal the prepared training-data inputs byte for byte, and the publisher standard model and the two margin-0.3 models reproduce their own training-run validation outputs word for word on 24 shared articles.

Length: every input is within 512 tokens (input_tokens). release tells which of Raunak's two releases an article came from: dedup (711 articles, 72–412 tokens) or full1024 (476 articles, 130–510 tokens).

Files

• merged.jsonl — one row per article: id (<story_id>:<side>), story_id, side, article_id, publisher, split, release, event_ids, title, original, nemotron_target, politics_original / politics_target (the hard-test's POLITICS-v2 scores of the original and the Nemotron target), input_tokens, and outputs = {model key: rewrite}.
• generations/<model key>.jsonl — one file per model with per-row metadata (terminated_with_eos, weights SHA, …).
• summary.json — counts, EOS rate and mean output words per model.

Caveats (from the hard-test README)

The POLITICS classifier selected these articles, so its scores on the originals sit at the ceiling by construction; compare models with each other and with the Nemotron targets rather than with zero. The left side is skewed toward NYT / Daily Kos / CNN.