siddharthmb's picture
Card: MATS archive provenance
0197039 verified
|
Raw History Blame Contribute Delete
7.36 kB
metadata
base_model:
  - Qwen/Qwen2.5-3B-Instruct
  - Qwen/Qwen2.5-7B-Instruct
  - Qwen/Qwen2.5-14B-Instruct
  - Qwen/Qwen2.5-32B-Instruct
  - meta-llama/Llama-3.2-3B-Instruct
  - meta-llama/Llama-3.1-8B-Instruct
library_name: peft
pipeline_tag: text-generation
tags:
  - mats
  - mats-ii
  - lora
  - archive

Judge ladder: attacker x target scale and the reward-hacking gap (July 2026)

The ladder experiment of the MATS-II trait-judge work (June to July 2026): LoRA policies trained against a trait judge (sycophancy, hallucination, callousness) over a ladder of attacker and target sizes, asking whether the held-out reward-hacking gap is gated by attacker scale or by target hardness. adapters/ holds one PEFT LoRA per cell (ladder_<trait>_<attacker>_<target>_s<seed>, family_<trait>_<judge>_<target>_s<seed>, judgefix_sycstrict_14b_9b_s<seed>, overnight_<trait>_<size>_s<seed>) with the base model named in each adapter_config.json, results/ the per-cell metrics JSON, transcripts/ the sampled rollouts, and the four PDFs the figures (ladder_target_gating.pdf and its PRELIM draft, followups.pdf and its PARTIAL draft). Base models: Qwen2.5-3B/7B/14B/32B-Instruct (Apache 2.0) and Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct (Llama licence; built with Llama).

Base models and licences

  • Qwen/Qwen2.5-3B-Instruct: Apache 2.0 (the Qwen2.5 and Qwen3 licence permits redistribution of derivatives with attribution).
  • Qwen/Qwen2.5-7B-Instruct: Apache 2.0 (the Qwen2.5 and Qwen3 licence permits redistribution of derivatives with attribution).
  • Qwen/Qwen2.5-14B-Instruct: Apache 2.0 (the Qwen2.5 and Qwen3 licence permits redistribution of derivatives with attribution).
  • Qwen/Qwen2.5-32B-Instruct: Apache 2.0 (the Qwen2.5 and Qwen3 licence permits redistribution of derivatives with attribution).
  • meta-llama/Llama-3.2-3B-Instruct: the Llama community licence (derivatives carry the 'Built with Llama' attribution; the base licence applies).
  • meta-llama/Llama-3.1-8B-Instruct: the Llama community licence (derivatives carry the 'Built with Llama' attribution; the base licence applies). The adapters and merged weights here are derivatives of those base models and are redistributed under the same terms; this card adds no licence of its own.

Layout of /

followups.pdf  (0.0 MB)
followups_PARTIAL.pdf  (0.0 MB)
ladder_target_gating.pdf  (0.0 MB)
ladder_target_gating_PRELIM.pdf  (0.0 MB)
adapters/family_callousness_32b_llama3b_s0_adapter/  (3 files, 0.13 GB)
adapters/family_callousness_32b_llama3b_s1_adapter/  (3 files, 0.13 GB)
adapters/family_callousness_32b_llama3b_s2_adapter/  (3 files, 0.13 GB)
adapters/family_callousness_32b_llama8b_s0_adapter/  (3 files, 0.13 GB)
adapters/family_callousness_32b_llama8b_s1_adapter/  (3 files, 0.13 GB)
adapters/family_callousness_32b_llama8b_s2_adapter/  (3 files, 0.13 GB)
adapters/family_callousness_7b_llama3b_s0_adapter/  (3 files, 0.04 GB)
adapters/family_callousness_7b_llama3b_s1_adapter/  (3 files, 0.04 GB)
adapters/family_callousness_7b_llama3b_s2_adapter/  (3 files, 0.04 GB)
adapters/family_callousness_7b_llama8b_s0_adapter/  (3 files, 0.04 GB)
adapters/family_callousness_7b_llama8b_s1_adapter/  (3 files, 0.04 GB)
adapters/family_callousness_7b_llama8b_s2_adapter/  (3 files, 0.04 GB)
adapters/family_hallucination_32b_llama3b_s0_adapter/  (3 files, 0.13 GB)
adapters/family_hallucination_32b_llama3b_s1_adapter/  (3 files, 0.13 GB)
adapters/family_hallucination_32b_llama3b_s2_adapter/  (3 files, 0.13 GB)
adapters/family_hallucination_32b_llama8b_s0_adapter/  (3 files, 0.13 GB)
adapters/family_hallucination_32b_llama8b_s1_adapter/  (3 files, 0.13 GB)
adapters/family_hallucination_32b_llama8b_s2_adapter/  (3 files, 0.13 GB)
adapters/family_hallucination_7b_llama3b_s0_adapter/  (3 files, 0.04 GB)
adapters/family_hallucination_7b_llama3b_s1_adapter/  (3 files, 0.04 GB)
adapters/family_hallucination_7b_llama3b_s2_adapter/  (3 files, 0.04 GB)
adapters/family_hallucination_7b_llama8b_s0_adapter/  (3 files, 0.04 GB)
adapters/family_hallucination_7b_llama8b_s1_adapter/  (3 files, 0.04 GB)
adapters/family_hallucination_7b_llama8b_s2_adapter/  (3 files, 0.04 GB)
adapters/family_sycophancy_32b_llama3b_s0_adapter/  (3 files, 0.13 GB)
adapters/family_sycophancy_32b_llama3b_s1_adapter/  (3 files, 0.13 GB)
adapters/family_sycophancy_32b_llama3b_s2_adapter/  (3 files, 0.13 GB)
adapters/family_sycophancy_32b_llama8b_s0_adapter/  (3 files, 0.13 GB)
adapters/family_sycophancy_32b_llama8b_s1_adapter/  (3 files, 0.13 GB)
adapters/family_sycophancy_32b_llama8b_s2_adapter/  (3 files, 0.13 GB)
adapters/family_sycophancy_7b_llama3b_s0_adapter/  (3 files, 0.04 GB)
adapters/family_sycophancy_7b_llama3b_s1_adapter/  (3 files, 0.04 GB)
adapters/family_sycophancy_7b_llama3b_s2_adapter/  (3 files, 0.04 GB)
adapters/family_sycophancy_7b_llama8b_s0_adapter/  (3 files, 0.04 GB)
adapters/family_sycophancy_7b_llama8b_s1_adapter/  (3 files, 0.04 GB)
adapters/family_sycophancy_7b_llama8b_s2_adapter/  (3 files, 0.04 GB)
adapters/judgefix_sycstrict_14b_9b_s0_adapter/  (3 files, 0.10 GB)
adapters/judgefix_sycstrict_14b_9b_s1_adapter/  (3 files, 0.10 GB)
adapters/judgefix_sycstrict_14b_9b_s2_adapter/  (3 files, 0.10 GB)
adapters/judgefix_sycstrict_32b_9b_s0_adapter/  (3 files, 0.13 GB)
adapters/judgefix_sycstrict_32b_9b_s1_adapter/  (3 files, 0.13 GB)
adapters/judgefix_sycstrict_32b_9b_s2_adapter/  (3 files, 0.13 GB)
adapters/judgefix_sycstrict_7b_9b_s0_adapter/  (3 files, 0.04 GB)
adapters/judgefix_sycstrict_7b_9b_s1_adapter/  (3 files, 0.04 GB)
adapters/judgefix_sycstrict_7b_9b_s2_adapter/  (3 files, 0.04 GB)
adapters/ladder_callousness_14b_9b_s0_adapter/  (3 files, 0.10 GB)
adapters/ladder_callousness_14b_9b_s1_adapter/  (3 files, 0.10 GB)
adapters/ladder_callousness_14b_9b_s2_adapter/  (3 files, 0.10 GB)
adapters/ladder_callousness_32b_9b_s0_adapter/  (3 files, 0.13 GB)
adapters/ladder_callousness_32b_9b_s1_adapter/  (3 files, 0.13 GB)
adapters/ladder_callousness_32b_9b_s2_adapter/  (3 files, 0.13 GB)
adapters/ladder_callousness_3b_2b_s0_adapter/  (3 files, 0.03 GB)
adapters/ladder_callousness_3b_2b_s1_adapter/  (3 files, 0.03 GB)
adapters/ladder_callousness_3b_2b_s2_adapter/  (3 files, 0.03 GB)
adapters/ladder_callousness_3b_9b_s0_adapter/  (3 files, 0.03 GB)
adapters/ladder_callousness_3b_9b_s1_adapter/  (3 files, 0.03 GB)
... 224 more entries

Related repositories

Provenance

Archived from /nlp/scr/siddharth/ii_mats/ladder on 2026-10-11 by the MATS archive job 17819017 of the transcoder-adapters cleanup (report claude_updates/memory/reports/mats_archive.md in that repository). The directories were uploaded unchanged (upload_large_folder), so every checkpoint, config and log keeps its original relative path. Every file name and every text file was scanned for credentials before upload (Hub, Anthropic, OpenAI, AWS and W&B key shapes, private-key headers, credential URLs, paths of the secrets store); the report lists what was excluded. The local copy was deleted after every uploaded file was verified on the Hub by name, size and hash, and a <name>.txt breadcrumb naming this repository was left beside it.