Steerling-8B-Instruct: Our Inherently Interpretable Instruct Model

Blue and black dialogue bubbles linked by visible networks of concept nodes

Author: Andreas Madsen, Founding Research Scientist
Published: November 17, 2026

We are releasing Steerling-8B-Instruct, the instruction-following variant of our interpretable base model, which we previously announced. This model retains all the interpretability properties the base model was trained to have, while also supporting multi-turn dialogues. It is now used in production on Clarity, our inherently interpretable AI platform. In this post, we cover how we trained the model and provide a checklist for others who want to perform instruction supervised fine-tuning (ISFT).

Steerling is a block-causal diffusion language model. The open-source model with the architecture closest to ours is LLaDA , which is also a diffusion-based language model. Although LLaDA uses a full-attention mask rather than Steerling’s block-causal mask, the instruction supervised fine-tuning (ISFT) pipeline should be the same. To ensure we could reliably instruction-tune diffusion-based models, we first reproduced LLaDA’s SFT results to verify that our pipeline worked, then applied it to our own Steerling model.

Reproducing LLaDA ISFT

As a first step, we reproduced the LLaDA-8B instruction supervised fine-tuning (ISFT) setup . The paper includes some important details, but we uncovered many additional details while reproducing the results. This section discusses our findings.

Data and Tokenization

Unfortunately, the LLaDA ISFT dataset is not publicly available. However, we found that allenai/tulu-3-sft-olmo-2-mixture, the same dataset used for ISFT of OLMo 2 , is sufficient to reproduce the English-language benchmark results. Based on the number of observations reported in the LLaDA paper , we trained for four epochs on this dataset.

To tokenize multi-turn dialogues, we apply the LLaDA template . This is essentially the Llama-3 template , but without a beginning-of-sequence (BOS) token, as diffusion models do not need one. The template looks like this:

<|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|>
<|start_header_id|>user<|end_header_id|>\n\n{user}<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>\n\n{assistant}<|eot_id|>

where <|start_header_id|>, <|end_header_id|>, and <|eot_id|> are special tokens.

LLaDA then unfolds the multi-turn dialogues: given pairs of user (ui\textbf{u}_i) and assistant (ai\textbf{a}_i) messages, an observation of {u1,a1,…,uI,aI}\{\textbf{u}_1, \textbf{a}_1, \dots, \textbf{u}_I, \textbf{a}_I\} becomes [{u1,a1},{u1,a1,u2,a2},⋯ ,{u1,a1,…,uI,aI}][\{\textbf{u}_1, \textbf{a}_1\}, \{\textbf{u}_1, \textbf{a}_1, \textbf{u}_2, \textbf{a}_2\}, \cdots, \{\textbf{u}_1, \textbf{a}_1, \dots, \textbf{u}_I, \textbf{a}_I\}]. Because LLaDA is not a causal model, unfolding is required for prompts containing only part of a dialogue to be in-distribution. For example, if the model had been trained without unfolding, providing only {u1,a1,t}\{\textbf{u}_1, \textbf{a}_{1,t}\} for t>0t > 0 (meaning a1,t\textbf{a}_{1,t} is partially masked) would be out-of-distribution, because the model would expect some of u2\textbf{u}_2 to be provided as well.

Checklist

When reproducing results, we often find that there is no special trick or magic hack that makes things work. It’s often just a matter of pursuing correctness at a level of detail that can seem obsessive. Sometimes, fixing one issue can make the model’s accuracy worse, but once you fix all the issues, things start to work. It can be hard to identify all these issues, but over time, we have developed a useful checklist.

Results

After following the checklist and fixing the benchmark issues, we were able to reproduce the LLaDA ISFT results , achieving nearly identical scores. Keep in mind that we used a different ISFT dataset and benchmark harness.

From paperReproduced (official model)ISFT on Tulu 3 SFT
MMLU65.564.7258.17
MMLU-Pro37.036.3140.11
ARC-C88.586.0187.37
GSM8K78.677.5676.27
Math26.626.6822.16
GPQA31.830.5831.70
HumanEval47.642.6840.24
MBPP34.240.2039.00
Avg.51.2350.5949.38

LLaDA-8B benchmark scores (%). Averages are unweighted means across all eight benchmarks. We did not include HellaSwag, as it is a likelihood benchmark that LLaDA adapted for multiple-choice generation, which is somewhat awkward. If it were included, the averages would be 53.82%, 53.57%, and 53.48%, respectively.

From LLaDA to Steerling

Training the Instruct version of Steerling is similar to training LLaDA. We use the same template, unfolding procedure, and general checklist to ensure correctness. However, while we were reproducing LLaDA’s results with the Tulu 3 dataset, the Dolci dataset was released, so we switched to it for Steerling. There are also some significant changes because Steerling uses block-causal diffusion and is inherently interpretable.

Block-causal diffusion and interpretability changes

Unlike LLaDA, Steerling uses block-causal diffusion. While LLaDA pads sequences with end-of-sequence tokens ([EOS]) to make them equal in length, we effectively pad only the last diffusion block that contains content (the finalizer block) with [EOS] tokens. Tokens after the finalizer block are never predicted. Therefore, in line with the checklist, these tokens should not contribute to the loss.

A significant difference between LLaDA and Steerling is that Steerling is inherently interpretable. We achieve this by annotating the dataset with concepts and applying a number of interpretability losses. This process is essentially the same in ISFT as in pretraining, so we will not go into detail about the annotations and losses here. You can read Scaling Interpretable Language Models to 8 Billion Parameters and The FineWeb Concept Atlas for those details.

ISFT does introduce some significant differences, though. Like the language model loss, the interpretability losses are applied only to the tokens we predict: the assistant tokens in the last message. However, we still want to support masking any input token to enable input-attribution explanations. We achieve this by applying diffusion masking to all input tokens, including both user and assistant message tokens, rather than just the predicted tokens in the last assistant message. The same trick works for LLaDA, and we found no performance penalty on average.

Benchmarks LLaDA-8B scores with and without input-token masking.
From paperReproduced using the official modelISFT on Tulu 3 SFTISFT on Tulu 3 SFT (with masking support)
MMLU65.564.7258.1760.12
MMLU-Pro37.036.3140.1140.01
ARC-C88.586.0187.3787.20
GSM8K78.677.5676.2773.31
Math26.626.6822.1622.82
GPQA31.830.5831.7031.70
HumanEval47.642.6840.2440.24
MBPP34.240.2039.0039.60
Avg.51.2350.5949.3849.38

LLaDA-8B benchmark scores (%), reproduced with and without input-token masking. Averages are unweighted means across all eight benchmarks. Masking all input tokens for attribution support does not materially affect benchmark performance.

Performance benchmarks

Here we compare Steerling’s benchmark results with those of other models. Steerling does not perform quite as well as some newer models, as discussed in Steerling-8B: The First Inherently Interpretable Language Model. This is because Steerling was pretrained with significantly fewer resources than other models, which are overtrained. When we account for this by plotting performance against pretraining FLOPs, Steerling performs as expected.

Performance and pre-training compute
MMLU · MMLU-Pro · ARC-C · GPQA · GSM8K · Math
All benchmarks: performance versus pre-training computeSteerling is highlighted in purple; comparison models are gold. Dashed lines show multiples of Steerling's compute. The x-axis is logarithmic.010203040506070801× (Steerling)2×4×8×16×4 × 10²²6 × 10²²1 × 10²³2 × 10²³4 × 10²³8 × 10²³1 × 10²⁴Approximate pre-training FLOPsAvg. performance (%)Steerling 8B: 33.60%; 6 benchmarks; 1.35T training tokens; 6.48 × 10²² FLOPsSteerling 8BLLaDA 8B: 54.67%; 6 benchmarks; 2.3T training tokens; 1.1 × 10²³ FLOPsLLaDA 8BOLMo 2 7B: 51.72%; 6 benchmarks; 4T training tokens; 1.92 × 10²³ FLOPsOLMo 2 7BLLaMA2 7B: 27.87%; 6 benchmarks; 2T training tokens; 9.6 × 10²² FLOPsLLaMA2 7BLLaMA3 8B: 55.42%; 6 benchmarks; 15T training tokens; 7.2 × 10²³ FLOPsLLaMA3 8BDeepSeek 7B: 37.39%; 6 benchmarks; 2T training tokens; 9.6 × 10²² FLOPsDeepSeek 7B

Average benchmark performance versus approximate pretraining FLOPs for Steerling-8B-Instruct and open SFT models of comparable size. Dashed lines mark multiples of Steerling’s compute. Compute is estimated as 6 × 8B parameters × pretraining tokens for all models. Scores for LLaDA, LLaMA2, LLaMA3, and DeepSeek are from the LLaDA paper . OLMo 2’s scores are from its paper , with ARC-C from Table 8 of Lin et al. . DeepSeek’s MMLU-Pro and GPQA scores are from the Open LLM Leaderboard .

Interpretability metrics

During post-training, we apply the same interpretability losses used during pretraining and mid-training, as described in Scaling Interpretable Language Models to 8 Billion Parameters. It is therefore natural to compare interpretability metrics across these training stages. Unfortunately, this is not straightforward. The validation metrics used during pretraining and mid-training are based on natural text continuations. However, the post-trained model is instruction-tuned and will not continue an input in the same way as a model from either earlier stage. To bridge this gap, we prompt the model with Continue the following text: {text}, then treat the assistant’s output in the same way as the output from the earlier models. This comparison is imperfect, but it allows us to use the same dataset.

Concept losses: When we evaluate the interpretability losses on a validation dataset, we find that the concept loss, which measures concept-classification accuracy, remains unchanged, while the independence loss increases.

Pre-trainedMid-trainedPost-trained
Concept losses across training stages: Concept loss Concept loss: Pre-trained 0.002, Mid-trained 0.002, Post-trained 0.002. Each metric groups the pre-trained, mid-trained, and post-trained models. All bars start at zero. 0 0.0005 0.001 0.0015 0.002 0.0025 Validation loss Concept loss — Pre-trained: 0.002 0.002 Concept loss — Mid-trained: 0.002 0.002 Concept loss — Post-trained: 0.002 0.002 Concept loss
Concept losses across training stages: Concept independence loss Concept independence loss: Pre-trained 1.907, Mid-trained 1.546, Post-trained 2.863. Each metric groups the pre-trained, mid-trained, and post-trained models. All bars start at zero. 0 0.5 1 1.5 2 2.5 3 3.5 Validation loss Concept independence loss — Pre-trained: 1.907 1.907 Concept independence loss — Mid-trained: 1.546 1.546 Concept independence loss — Post-trained: 2.863 2.863 Concept independenceloss
Validation concept losses across training stages. The two panels use separate scales.

Concept contribution: Despite the increase in independence loss, when we examine the output contributions from known concepts, discovered concepts, and the residual stream, we find that the contribution from known concepts has increased, while the residual contribution has decreased. Overall, this suggests that post-training has improved our concept attribution and ability to steer the model.

Pre-trainedMid-trainedPost-trained
Concept contributions across training stages Known concepts: Pre-trained 48.9%, Mid-trained 48.9%, Post-trained 50.3%. Discovered concepts: Pre-trained 38.6%, Mid-trained 38.6%, Post-trained 39.0%. Residual: Pre-trained 12.5%, Mid-trained 12.5%, Post-trained 10.7%. Each metric groups the pre-trained, mid-trained, and post-trained models. All bars start at zero. 0% 10% 20% 30% 40% 50% 60% Share of logit magnitude (%) Known concepts — Pre-trained: 48.9% 48.9% Known concepts — Mid-trained: 48.9% 48.9% Known concepts — Post-trained: 50.3% 50.3% Known concepts Discovered concepts — Pre-trained: 38.6% 38.6% Discovered concepts — Mid-trained: 38.6% 38.6% Discovered concepts — Post-trained: 39.0% 39.0% Discovered concepts Residual — Pre-trained: 12.5% 12.5% Residual — Mid-trained: 12.5% 12.5% Residual — Post-trained: 10.7% 10.7% Residual
Token-level logit contributions from known concepts, discovered concepts, and the residual stream.

Known concept alignment: Finally, we measure concept alignment by comparing concept attributions with assessments from an LLM judge. We find that concept alignment has also improved.

Pre-trainedMid-trainedPost-trained
Known concept alignment across training stages Known concept alignment: Pre-trained 3.730, Mid-trained 3.770, Post-trained 3.810. Each metric groups the pre-trained, mid-trained, and post-trained models. All bars start at zero. 0 1 2 3 4 5 Alignment score Known concept alignment — Pre-trained: 3.730 3.730 Known concept alignment — Mid-trained: 3.770 3.770 Known concept alignment — Post-trained: 3.810 3.810 Known concept alignment
Known concept alignment with an LLM judge across training stages.

Overall, post-training yields mostly identical or slightly better results on interpretability metrics. One possible explanation is that concepts are less ambiguous in an instruction-following context than in a pure text-continuation context.

Release

Our Steerling-8B-Instruct model, the first inherently interpretable large language model, is now in production on Clarity. We have also made it available for anyone to use on Hugging Face at guidelabs/steerling-8b-instruct.

References

  1. S. Nie, F. Zhu, Z. You et al.. Large language diffusion models . The thirty-ninth annual conference on neural information processing systems, 2025.
  2. , P. Walsh, L. Soldaini et al.. 2 olmo 2 furious . 2024.
  3. A. Grattafiori, A. Dubey, A. Jauhri et al.. The llama 3 herd of models . arXiv, 2024.
  4. J. Austin, A. Odena, M. Nye et al.. Program synthesis with large language models . , pp. 1–34, 2021.
  5. T. Olmo, A. Ettinger, A. Bertsch et al.. Olmo 3 . 2025.
  6. P. Lin, R. Balasubramanian, F. Liu et al.. Efficient model development through fine-tuning transfer . Proceedings of the 2025 conference on empirical methods in natural language processing, pp. 2617–2636. Association for Computational Linguistics, 2025.
  7. Open LLM Leaderboard: Evaluation results for DeepSeek-LLM-7B-Chat . Hugging Face, 2024.