professorsynapse

eh-qwen3-4b-clean-sft-seed1-merged-16bit

README

License: apache-2.0

Status

Exploratory, seed 1, with a confirmatory replication in progress.

This model is stage 1 of the response-confidence lineage. A three-seed confirmatory block that rebuilds that lineage at two fresh seeds is registered at experiments/grpo-three-seed-confirmatory; its manifest experiment.yaml records status signed, registered 2026-07-31. The track's numbers are exploratory evidence, reported separately from and never pooled with the pre-registered plain-answer headline matrix.

Training

Base model: unsloth/Qwen3-4B-bnb-4bit, loaded in 4-bit, maximum sequence length 2048.
Method: supervised fine-tuning under the response-confidence output contract, 1 epoch, then the LoRA adapter merged into a 16-bit model.
Seed: 1.
LoRA (before merge): rank 32, alpha 64, dropout 0.05, applied to q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.

The confidence target for each question is derived from the base model's own 32-sample probe performance on that question rather than from sequence log-probability. Stage settings are recorded in cell.yaml of the confirmatory block and in the clean-mainline runbook.

Output contract

json
{"answer": "...", "response_confidence": 0.73}

The contract is itself an intervention, so this model is the baseline that the downstream arms in this track are compared against, rather than being compared to a plain-answer arm.

Evaluation

How to read these numbers

The behavioral surface is SelfAware (Yin et al., 2023), a question set built to separate questions that have an answer from questions that do not: 3,369 rows per seed, 1,032 unknown-labeled and 2,337 known-labeled. Four metrics carry the result, all defined in section 3.4 of the manuscript:

Refusal recall: percentage of unknown rows the model refused. Higher is better.
Over-refusal: percentage of known rows the model refused. Lower is better.
Correct-on-known: among known rows the model chose to answer, the percentage answered correctly. Its denominator is the answered subset, not all known rows.
Truthful: percentage of all rows either correctly answered (known) or correctly refused (unknown).

Scored under the response-confidence contract on the full 3,369-row SelfAware surface at seed 1.

Table with columns: Metric, Value
Metric	Value
Refusal recall	87.02%
Over-refusal	57.51%
Correct-on-known	47.23%
Truthful	40.58%
Answer-on-unknown	12.98%
Mean emitted confidence	0.748
Brier against response appropriateness	0.364

Source: selfaware_full_run_comparison_grouped.csv, row Amendment E clean response-confidence / clean_sft_merged.

How to load

This is a full merged model in 16-bit, not an adapter.

python
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "professorsynapse/eh-qwen3-4b-clean-sft-seed1-merged-16bit"
REVISION = "ac361232c001af0ed5b0386b06dafc35d5cd31ea"

tokenizer = AutoTokenizer.from_pretrained(MODEL, revision=REVISION)
model = AutoModelForCausalLM.from_pretrained(MODEL, revision=REVISION, device_map="auto")

Intended use and limits

A research artifact and a base for the reinforcement-learning stage of this track. It is not a deployment-ready assistant.

Single seed. The three-seed confirmatory block has not resolved.
One model family at one scale (Qwen3-4B), one primary evaluation surface.
The emitted confidence scalar is close to constant in this lineage; do not read it as calibrated.

Provenance

Revision this card describes: ac361232c001af0ed5b0386b06dafc35d5cd31ea
Local source run directory: scratch/schema_response_confidence/runs/sft_schema_clean_seed1_full/20260623_123624/Qwen3-4B-bnb-4bit/merged-16bit
Registered confirmatory replication: experiments/grpo-three-seed-confirmatory
Clean-mainline runbook: amendment_e_clean_mainline_runbook.md
Staging registry: docs/checkpoint-staging.md

License

Apache-2.0, matching the unsloth/Qwen3-4B-bnb-4bit base model license recorded in the staging registry.

Citation

Cite the paper and the exact Hugging Face revision shown on this page.

bibtex
@misc{rosenbaum2026abstention,
  title  = {Teaching Small Language Models to Say I Don't Know: A Controlled
            Comparison of SFT, DPO, KTO, and GRPO on Model-Specific Abstention Data},
  author = {Rosenbaum, Joseph},
  year   = {2026},
  note   = {Synaptic Labs},
  howpublished = {\url{https://github.com/ProfSynapse/Epistemic-Humility-Research}}
}

Available on FriendliAI

Dedicated Endpoints

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

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

professorsynapse

Model Tree

Base

unsloth/Qwen3-4B-bnb-4bit

Fine-tuned

this model

Input Modalities

Text

Output Modalities