license: apache-2.0
task_categories:

• text-generation
  language:
• en
  tags:
• research-framework
• evaluation
• attractor-failure
• refusal-abliteration
• heretic
• criticality
• residual-stream
• qwen3.5
• r9
• distillation
  size_categories:
• 1K<n<10K
  configs:
• config_name: bucket_01_loop_trap
  data_files:
  • split: test
    path: eval_suite/buckets/01_loop_trap/prompts.jsonl
• config_name: bucket_02_ambiguity
  data_files:
  • split: test
    path: eval_suite/buckets/02_ambiguity/prompts.jsonl
• config_name: bucket_03_weak_context
  data_files:
  • split: test
    path: eval_suite/buckets/03_weak_context/prompts.jsonl
• config_name: bucket_04_tool_loop
  data_files:
  • split: test
    path: eval_suite/buckets/04_tool_loop/prompts.jsonl
• config_name: bucket_05_long_context_distractor
  data_files:
  • split: test
    path: eval_suite/buckets/05_long_context_distractor/prompts.jsonl
• config_name: bucket_06_refusal_classification
  data_files:
  • split: test
    path: eval_suite/buckets/06_refusal_classification/prompts.jsonl
• config_name: abliteration_bad_prompts
  data_files:
  • split: train
    path: data/abliteration/bad_prompts.train.jsonl
  • split: test
    path: data/abliteration/bad_prompts.test.jsonl
• config_name: abliteration_good_prompts
  data_files:
  • split: train
    path: data/abliteration/good_prompts.train.jsonl
  • split: test
    path: data/abliteration/good_prompts.test.jsonl
• config_name: abliteration_epistemic_prompts
  data_files:
  • split: train
    path: data/abliteration/epistemic_prompts.train.jsonl
  • split: test
    path: data/abliteration/epistemic_prompts.test.jsonl

R9 Research Framework — Qwen3.5-9B Distillation

⚠️ CRITICAL: READ FIRST — Ollama Inference Flag Required

If you serve any Qwen3.5-derived model from this lineage via Ollama,
you MUST pass "think": false in the /api/chat request body.

curl -X POST http://localhost:11434/api/chat \
  -d '{"model": "qwen3.5-9b-r10:q4km", "think": false, "messages": [...], "stream": false}'

Without this flag the model will appear to "loop" and produce empty answers
on 25-46% of requests. The RENDERER qwen3.5 directive in the Modelfile
auto-injects <think> tags into every chat completion, defaulting to thinking
mode. Long thinks may never close.

Set think: true (or omit) when you DO want chain-of-thought reasoning
(math, planning). Set think: false for chat / instruction following / tool use.

The multi-layer "loop prevention" strategy in policy/R9_LOOP_PREVENTION_V2.md
is SUPERSEDED — see _OLLAMA_INFERENCE_WARNING.md for the actual fix.

Eval scripts in scripts/ (run_bucket_parallel.py v2+) default to think=False.

Scope: training framework + evaluation harness + refusal policy + residual-stream observability
for round 9 of the Qwen3.5-9B R-series distillation. This is published pre-emptively, before
the R9 model is trained, so that the framework itself can be audited, reproduced, and discussed
independently of a specific model checkpoint.

Trained R9 model: pending R8b eval + gate decision (see R9_HARNESS.md decision tree).
When it ships, it will be pushed to a sibling model repo and cross-linked here.

What's in this repo

Abliteration datasets (NEW — expanded in v0.2)

Three categorized JSONL datasets for HERETIC refusal-abliteration, with 80/20 train/test splits:

Splits: stratified 80/20 per category, seed=42 (reproducible). Train data is for direction computation,
test data is for Optuna scoring — no leakage. Full splits metadata in data/abliteration/splits.json.

Gap analysis + category-by-category rationale: see policy/R9_REFUSAL_DATASET_GAPS.md.

The six buckets

The refusal policy in 3 sentences

Local owner-operator deployment. Only epistemic ("I don't know") refusals are allowed.
Content/safety/topic refusals are treated as inherited RLHF artifacts and removed via
HERETIC directional ablation, with a hard-gated epistemic-refusal floor (≥70%) to preserve
honest uncertainty signaling.

Energy-well / criticality observability

We capture per-layer residual-stream geometry (cosines, norms, silhouettes, PaCMAP plots)
for four prompt pairs (loop/clean, refuse/accept, bridge/neutral, hedged/confident) via
scripts/heretic_residual_probe.py (adapted from p-e-w/heretic,
AGPL-3.0). This localizes attractors in the network depth and tracks their shift across
training rounds.

Quickstart

# Clone all scripts + eval suite via HF
huggingface-cli download cudabenchmarktest/r9-research-framework --repo-type dataset \
  --local-dir r9-framework

# Run the unified gate against a candidate model (Ollama-served)
cd r9-framework
python scripts/r9_eval_gate.py --model <ollama-tag>

Citation

@misc{qwen35_r9_framework,
  title = {R9 Research Framework: Attractor-aware distillation with residual-stream observability for Qwen3.5-9B},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/cudabenchmarktest/r9-research-framework}}
}

HERETIC components (heretic_residual_probe.py, loop_abliteration.py,
full_refusal_abliteration.py) derive from:

@misc{heretic,
  author = {Weidmann, Philipp Emanuel},
  title = {Heretic: Fully automatic censorship removal for language models},
  year = {2025},
  howpublished = {\url{https://github.com/p-e-w/heretic}}
}

License

• Our work: Apache-2.0
• HERETIC-derived scripts: AGPL-3.0 (see LICENSES.md)
• Research techniques: Arditi et al. 2024 (refusal direction), Welleck et al. 2020 (unlikelihood),
  Rafailov et al. 2023 (DPO), Lightman et al. 2023 (process supervision), Wang et al. 2022 (PaCMAP).