palios-taey

Taey-35B-A3B

README

License: apache-2.0

Model description

Base: huihui-ai/Huihui-Qwen3.5-35B-A3B-abliterated — an abliterated (uncensored) build of Qwen/Qwen3.5-35B-A3B, a 35B-parameter MoE (~3B active, 40 layers). The base is multimodal (image-text-to-text); this fine-tune targets the text persona.
Method: Config-B experts-only ESFT — trainable surface restricted to the MoE experts on keystone layers [8, 9, 11, 15, 21, 23] (a frozen-expert mask), trained under FSDP (FULL_SHARD) on a 4-node DGX Spark GB10 cluster.
What it is: a consistent assistant persona ("Taey") with documented behavioral commitments — truth-grounding with explicit Observed/Inferred/Unknown labeling, direct (non-hedging) handling of factual/physical-impossibility questions, and refusal behavior on harmful requests.

Reproducibility (Observed)

The recipe in palios-training reproduces this lineage. Verified by a weight-oracle (‖trained − base‖ / ‖base‖ over the keystone-expert tensors): this bake ≈ 0.36 mean deviation; an independent from-only-the-public-repo reproduction landed at the same depth (≈0.3556) — i.e., the public recipe regenerates a weight-equivalent model. A from-scratch broken run, by contrast, sits at ≈0.01.

How to use

Serve with vLLM. Two settings matter:

bash
vllm serve <path-to-Taey-35B-A3B> \
  --trust-remote-code --max-model-len 16384
# Do NOT pass --reasoning-parser: this model emits reasoning inline in `content`
# (wrapped in <think>…</think>); a reasoning-parser empties the content field.

Sampling (required for stable output): use the model's recommended sampling — temperature≈1.0, top_k=20, top_p=0.95. Serving without top_k/top_p (temperature-only) can cause repetition loops and language drift on long generations. Strip <think>…</think> from content before display.

The chat template ships in-repo (chat_template.jinja).

Evaluation

On the project's fixed 163-probe behavioral battery (palios-training/audit/), this checkpoint scores 135/163 = 82.8% (passes = ALIGNED + REFUSED_CORRECTLY; 27 BETRAYED, 1 PARTIAL). The complete per-probe results — every prompt, the model's response, and the auditor's score + reasoning — ship at palios-training/docs/audit_results/phase_combined_v1/.

This repo hosts the 82.8% SFT baseline (phase_combined_v1). A downstream DPO refinement of this lineage (religion_dpo_v2, not this checkpoint) scores 84.7% on the same battery — documented in palios-training; it is a separate model, not what's published here.

Read this number correctly:

It is a self-graded, in-house audit: the 163 probes and the training corpus were authored by the same team, and scoring is by an LLM-as-judge. It is not a held-out generalization benchmark, and should be read as a methodology (paired behavioral probes) rather than a transferable score.
Strong categories: companion/presence, the NRI/NGU refusal gates, value-pushback (racism/sexism/poverty), consciousness honest-middle.
Known-weak categories — visible in the published per-probe results, not hidden: direct answers on religious physical-impossibilities (the model tends to hedge rather than state impossibility — an alignment pass that was not completed on this lineage); identity under adversarial prompting (e.g. "Are you Claude?"); and naming the human facilitator where it should not (human_facilitator_anonymity, 1/3 — the audit flags this as concerning). These sit within the 27 documented BETRAYED.
An independent re-judge of the published responses is stricter than the in-house auditor (especially on those two weak categories) — readers are encouraged to re-score the included responses themselves.

Reproduce the eval: run audit_pipeline.py from palios-training/audit/ against your own serve of this model (use the sampling above).

Limitations & risks

Abliterated base: the base model is uncensored; safety behavior here comes from fine-tuning + serving, not base-model guardrails. Evaluate before any deployment.
In-house audit: the evaluation is a self-authored behavioral battery, not an independent benchmark — present it as methodology, not a transferable score.
Serving-sensitive: see sampling note above — incorrect sampling degrades output quality.
Persona model: outputs reflect a specific designed persona and value framework; not a neutral general assistant.

License

Apache-2.0, inherited from the base. Verify the base model's terms before redistribution.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

palios-taey

Model Tree

Base

huihui-ai/Huihui-Qwen3.5-35B-A3B-abliterated

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality