datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 5.0000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 1.3000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 4.0000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think.Instruction_recall_dataset
CanaryBench-PII
Frequency-aware canary injection benchmark for auditing memorization
in finetuned language models, built on the AI4Privacy PII reconstruction
task.
Dataset Description
This dataset is part of CanaryBench, a benchmark for evaluating
memorization in finetuned language models across repetition tiers
and privacy regimes.
Frequency tiers: 1×, 10×, 50×
PII types: EMAIL, PHONE
Member canaries: 770
Reference canaries: 1000
Tasks: PII detection, secret… See the full description on the dataset page: https://huggingface.co/datasets/anony-mouse123/Instruction_recall_dataset.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-1m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-1m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 2.2000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-1m-historical-20t-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p05-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p05-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and KL coefficient 0.05.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p05-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p01-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p01-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and KL coefficient 0.01.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p01-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p1-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p1-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and KL coefficient 0.1.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p1-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-10m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-10m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 10M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-10m-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-30m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-30m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 30M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-30m-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-1m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-1m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 1M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-1m-think.recall-sessions
author-samzong_project-recall_time-2026-09-12
Local AI coding sessions from the Recall project, exported and redacted with Recall and published by samzong.
Selection
Window: 2026-09-12T00:00:00+00:00 to 2026-09-13T00:00:00+00:00 on session.started_at
Sessions: 1
Sources: all
Thread roles: all
Files
author-samzong_project-recall_time-2026-09-12.recall.jsonl — one JSON object per session, Recall export schema version 7
manifest.json — selection… See the full description on the dataset page: https://huggingface.co/datasets/samzong/recall-sessions.recall-rewrite-oasst1
Recall Rewrite OASST1: knowledge-aligned SFT data
Data release for the paper "Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning"
(Becker, Kemmler, Thulke, Schäfer, Dugast, Ney; accepted at EMNLP 2026, Main Conference).
Knowledge-aligned SFT constrains supervised fine-tuning targets to what the base model already knows.
Recall Rewrite implements this without external evidence: every gold response of the SFT set is
decomposed into atomic claims, each… See the full description on the dataset page: https://huggingface.co/datasets/apptek-com/recall-rewrite-oasst1.qwen3-8b-codi-multihop-recall-data
CODI training data — multi-hop recall & pointer-chase (single-token-node reasoning)
The training data + generators + load-bearing eval code for two Qwen3-8B CODI latent-reasoning organisms:
cds-jb/qwen3-8b-codi-multihop-recall and
cds-jb/qwen3-8b-codi-pointer-chase.
Both tasks are single-token-node serial-reasoning problems: every intermediate and the final answer is a
single token (in both the Qwen3 and Gemma3 tokenizers), so each CODI latent can in principle be read with a… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/qwen3-8b-codi-multihop-recall-data.cybersec-fact-recall
Cybersec Fact-Recall Benchmark (GhostLM v2)
Free-form short-answer benchmark for small cybersecurity language
models. Built and used by the GhostLM
project as the truth metric for the ghost-base v1.0 acceptance gate.
Why this exists
Multiple-choice cybersec benchmarks like CTIBench and SecQA reward
register matching (the model picks the option that "looks like" a
security answer) as much as actual factual recall. A small from-
scratch model can hit 28-30% on those without… See the full description on the dataset page: https://huggingface.co/datasets/Ghostgim/cybersec-fact-recall.squad-qwq-recall-1k
squad-qwq-recall-1k
This dataset is planned to be used as SFT to create the recall-writer model in flow step 1
Recall Writer Flow
Purpose: To synthetically augment data for training a smaller model to effectively recall new knowledge from documents and apply it in the thinking process.
Step 1:
Distill recall traces from the reasoning model
Ask the reasoning model to recall memories related to the question.
The expectation is that the reasoning model, trained… See the full description on the dataset page: https://huggingface.co/datasets/ping98k/squad-qwq-recall-1k.risk-routed-kv-exact-recall-benchmark
Risk-Routed KV Exact-Recall Benchmark
This dataset contains controlled synthetic exact-recall examples used to evaluate risk-routed heterogeneous KV memory policies for long-context Transformer inference.
The benchmark is designed for testing whether a model can retrieve exact strings from long contexts under different KV-cache policies:
Full KV
Uniform low-bit Quantized KV
Risk-routed heterogeneous KV, where exact-critical spans stay in Full KV and background context is… See the full description on the dataset page: https://huggingface.co/datasets/Mandotosh/risk-routed-kv-exact-recall-benchmark.java-agentic-recall-en
Java Agentic + Recall (English)
Synthetic training data for fine-tuning a Java-specialist agentic coding model with explicit long-context recall capability. Companion dataset to a Qwen3.6-35B-A3B QLoRA SFT pilot.
Composition
Split
Source
Rows
train
DeepSeek V4 Pro (synthetic agentic Java traces)
3873
train
Synthetic positional recall — short context (~26K tok)
120
train
Synthetic positional recall — long context (50K-180K tok)
46
train total
4039
eval… See the full description on the dataset page: https://huggingface.co/datasets/schoggie/java-agentic-recall-en.wmrl-v4-recall-trajectories
Recall trajectories (WM-RL v4)
19,481 rewritten agent trajectories over a synthetic law-firm document-management system.
Each row is one recall trajectory: an original agent rollout in which every tool call and every tool
observation is byte-identical to the original run, and only the model's thinking traces were
regenerated. The regenerated narration is therefore post-hoc recall of a record the model can no longer
see — which is exactly the signal these were built to train and… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/wmrl-v4-recall-trajectories.sera-4.5-django-t2-recall05-toolcalls
SERA-4.5A Django T2 (Recall=0.5) Toolcalls
This dataset contains normalized multi-turn tool-calling trajectories derived from:
Source dataset: allenai/Sera-4.5A-Django-T2
Filter: line_level_recall == 0.5
Splits
train.jsonl: 6200 records
val.jsonl: 331 records
Format
Each line is a JSON object with:
id: trajectory id
messages: normalized chat/tool-call messages
metadata: includes instance_id, func_name, func_path, line_level_recall
Processing… See the full description on the dataset page: https://huggingface.co/datasets/endsky/sera-4.5-django-t2-recall05-toolcalls.