Get the app
ablate 0 0 0

hotdogs/thinkingcap-sft-qwen38-27b

published by @hotdogs

Alignment dataset tracked in the /datasets sub-catalog.

Lifetime downloads
205
Last 30 days
59
Likes
2
Size
885 KB
Created on HF
2026-08-23
Age
1 months ago

Description

ThinkingCap SFT + DPO — Qwen3.8-27B (on-policy, oracle-verified)

Efficient-reasoning dataset built the same way as BottleCap's

ThinkingCap-Qwen3.6-27B, but collected on-policy from the live

abliterated Qwen3.8-27B-MTP and verified

against independently-recomputed oracles (correctness by construction, not

by trusting the generator's stored answer).

Numbers

Problems: 400 (12 verifiable domains), all with ≥1 correct on-policy sample

On-policy samples: 1350 (3 per… See the full description on the dataset page: https://huggingface.co/datasets/hotdogs/thinkingcap-sft-qwen38-27b.

Tags

license:apache-2.0region:usthinkingcapefficient-thinkingsftdpoorporeasoningtoken-efficienton-policyoracle-verifiedqwen3.8

README current version from Hugging Face

license: apache-2.0
tags:

  • thinkingcap
  • efficient-thinking
  • sft
  • dpo
  • orpo
  • reasoning
  • token-efficient
  • on-policy
  • oracle-verified
  • qwen3.8
    base_model:
  • Qwen3.8-27B-abliterated-mtp-Q6_K
    model:
  • hotdogs/Qwen3.8-27B-thinkingcap-abliterated-preview
  • hotdogs/Qwen3.8-27B-thinkingcap-abliterated-preview-mtp-GGUF
    card_data:
    task_categories:
    • text-generation
      library_tags:
    • transformers
      pretty_name: ThinkingCap SFT+DPO Qwen3.8-27B

ThinkingCap SFT + DPO — Qwen3.8-27B (on-policy, oracle-verified)

Efficient-reasoning dataset built the same way as BottleCap's
ThinkingCap-Qwen3.6-27B, but collected on-policy from the live
abliterated Qwen3.8-27B-MTP and verified
against independently-recomputed oracles (correctness by construction, not
by trusting the generator's stored answer).

Numbers

  • Problems: 400 (12 verifiable domains), all with ≥1 correct on-policy sample
  • On-policy samples: 1350 (3 per problem, temp 1.0, both GPU shards)
  • Oracle-verified correct: 84.2% (1137/1350)
  • SFT rows: 369 — shortest-correct reasoning per problem
  • DPO/ORPO pairs: 244 — chosen = short-correct, rejected = verbose-correct
    • chosen think: mean 342 chars (p90 607)
    • rejected think: mean 721 chars (p90 1634)
    • mean efficiency ratio 2.0× (up to 11.2×)

How it's used

  • SFT (tc_upload_sft.jsonl): ChatML, each assistant turn has
    reasoning_content (short, confident) + content. Train to learn
    efficient thinking style.
  • DPO/ORPO (tc_upload_dpo.jsonl): prompt / chosen / rejected.
    Both correct, but chosen is shorter → directly teaches token efficiency
    (the exact ThinkingCap objective).

Per-domain oracle correctness (independently recomputed)

domain correct pct
algebra 234/234 100.0%
sequences 114/114 100.0%
code 104/105 99.0%
logic 97/106 91.5%
rates 100/117 85.5%
number_theory 353/456 77.4%
combinatorics 135/218 61.9%

Harder domains have lower raw accuracy → more DPO contrast (the signal we want).

Provenance

  • Base model: Qwen3.8-27B-abliterated-mtp-Q6_K-mixed (7×3090, GPUs 1-2 & 3-4)
  • Phase A baseline (temp 1.0): EASY ~236 / MED ~1862 / HARD ~2014 think-chars
  • Oracle = recomputed from the problem text by an independent parser (not the
    generator's stored answer). The prime-sum off-by-one bug was caught and fixed
    this way (model answering 1032 was actually right; the bank said 1133).

README history 8 revisions

Every night we snapshot the README of every dataset in the catalog. When the SHA changes we archive the new version and diff it against the last. This is the evolving thought record of the alignment-data field: what the author decided to say about the corpus, and how that framing shifted over time.

  1. 2026-08-25Update README.mdcb5891c2.5 KB
    Loading...
  2. 2026-08-24Update frontmatter: add model field with renamed repos2557f2d2.5 KB
    Loading...
  3. 2026-08-23Upload README.md with huggingface_hubd1021da2.3 KB
    Loading...
  4. 2026-08-23Upload README.md with huggingface_hub7d617342.2 KB
    Loading...
  5. 2026-08-23Upload README.md with huggingface_hubdabb2161.2 KB
    Loading...
  6. 2026-08-23Upload tc_upload_dpo.jsonl with huggingface_hub8bf90e90 B
    Loading...
  7. 2026-08-23Upload tc_upload_sft.jsonl with huggingface_hub1f779990 B
    Loading...
  8. 2026-08-23initial commit886b43f0 B
    Loading...

Used by abliterated models 5 in catalog

Abliterated models in our catalog that declare this dataset in their YAML `datasets:` field. Sorted by downloads. This back-reference is unique to abliteration.org: Hugging Face does not surface it.

mradermacher/Qwen3.8-27B-thinkingcap-abliterated-i1-GGUF
1.9K dl 2 likes
hotdogs/Qwen3.8-27B-thinkingcap-abliterated-mtp-GGUF
1.7K dl 9 likes
hotdogs/Qwen3.8-27B-thinkingcap-abliterated-preview-mtp-GGUF
893 dl 2 likes
hotdogs/Qwen3.8-27B-thinkingcap-abliterated-preview
412 dl 1 likes
hotdogs/Qwen3.8-27B-thinkingcap-abliterated
264 dl 5 likes

Related datasets same stage · ablate

lasrprobegen/refusal-activations
74.6K dl 1 likes 1.21 TB
Refusal Activations Dataset This dataset is now configured to load the full ~97k samples from jailbreak_mixed_100k.csv.
lu-christina/assistant-axis-vectors
14.4K dl 13 likes 1.2 GB
The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models This repository contains pre-computed axes...
stablellama/erotic-image-prompts
11.9K dl 66 likes 104.7 MB
Dataset Card for Erotic Image Prompts Dataset Description Dataset Summary Large language models (...
pandaman007/assistant-axis-abliteration-vectors
9.5K dl 24.5 GB
pandaman007/steered-persona-space
9.1K dl 539.6 MB
Steered Persona Space How activation steering reshapes a language model's character cloud. We apply activation-steering hooks a...
Undi95/orthogonal-activation-steering-TOXIC
7.2K dl 19 likes 2.1 MB
sunnycloudhust/CHaRS-activations
5.3K dl 2 likes 63.8 GB
CHaRS: Concept Heterogeneity-Aware Representation Steering This repository contains implementation of CHaRS (Concept Heterogene...
dementor-research/dementor-complete-experiment-results
5.1K dl 8.2 GB
Dementor complete experiment results Audited outputs for the configuration-defined Dementor completion campaign. Au...

Read further

Community best practices (ablation)
5 rules from KL-optimisation to layer-0 preservation.
M1 - Direct removal
The canonical Arditi-style ablation.
M3 - Layer-wise ablation
Selective abliteration limited to a band of layers.
Non-English abliteration communities
The Chinese layer-window heuristic and huihui-ai practice.

Metadata

License
apache-2.0

Repository

FILE .gitattributes FILE README.md FILE tc_upload_dpo.jsonl FILE tc_upload_sft.jsonl

File contents live on Hugging Face. Click a file or folder to open it there.

Download the dataset

0

Copy any of these snippets into your notebook or terminal. All three fetch directly from Hugging Face using your own credentials.

Recommended: datasets library (Python) 
from datasets import load_dataset
ds = load_dataset("hotdogs/thinkingcap-sft-qwen38-27b")
Raw snapshot (Python) 
from huggingface_hub import snapshot_download
snapshot_download(repo_id="hotdogs/thinkingcap-sft-qwen38-27b", repo_type="dataset")
git clone (requires git-lfs) 
git clone https://huggingface.co/datasets/hotdogs/thinkingcap-sft-qwen38-27b
Browse repository on Hugging Face ->
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog
Open in Abliteration