HRM-Text

A 1 B-parameter language model checkpoint built on the Hierarchical Reasoning Model (HRM) architecture, trained by Sapient Intelligence from scratch on structured public datasets.

HRM is a dual-timescale recurrent architecture: two Transformer modules (H = high-level / slow, L = low-level / fast) iterate over the same input embeddings for H_cycles × (L_cycles + 1) steps, with additive state injection (z_L + z_H). This gives effectively unbounded compute depth at bounded parameter count.

Disclaimer

This is a pre-alignment model checkpoint, not a chat or instruction-following assistant. It is pre-trained on a PrefixLM objective with condition prefix tokens and has not been multi-turn dialogue tuned, long-context adapted, instruction-tuned, RLHF-trained, or otherwise aligned for assistant-style use. If you want to use HRM-Text like a chat model, you would need to perform further alignment, such as SFT and/or RL, on task-specific data. This checkpoint is meant to serve as a starting point, not a finished assistant.

Practical guidance for prompting the raw checkpoint:

• NLP tasks (classification, extraction, structured output, short-form QA): use the direct condition with 2–8 few-shot in-context examples. direct + few-shot is the strongest zero-extra-training setup we have measured; pure zero-shot is noticeably weaker.
• Reasoning / math / open-ended generation: use the composite condition synth,cot. This is one composite prefix, not two alternatives — at tokenization time the comma-separated tags are mapped to their prefix tokens and concatenated, in order, into a single prefix block. So synth,cot produces the two-token prefix (synth first, then cot), wrapped in the usual … `` envelope. Under this composite the model exhibits some chain-of-thought / instruct-like behavior — enough to answer many zero-shot math and reasoning prompts in a step-by-step style — but quality is uneven and below an instruction-tuned model of comparable size. Treat this "instruct" ability as a side effect of the pre-training mix, not a guaranteed capability.

The four single condition tags and their assigned tokenizer special tokens (token names are legacy implementation details; you can compose any subset, comma-separated, in the order you want them emitted):

• direct → `` — direct answer, no CoT
• cot → `` — chain-of-thought
• noisy → `` — noisy / web-crawl style
• synth → `` — synthetic / curated style

Requirements

Requires transformers >= 5.9.0, which ships native support for the hrm_text model class:

pip install --upgrade "transformers>=5.9.0"

Model details

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "sapientinc/HRM-Text-1B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
).cuda().eval()

# synth,cot composite — reasoning / CoT style (see Disclaimer for other modes)
condition = ""
prompt = f"{condition}Explain why the sky is blue."

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
# Mark the prompt as a single bidirectional prefix block — see "PrefixLM mask" below.
inputs["token_type_ids"] = torch.ones_like(inputs["input_ids"])

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=False))

PrefixLM mask — pass token_type_ids

HRM-Text was pre-trained with a PrefixLM mask: prompt tokens attend bidirectionally to each other, response tokens attend causally. To match the training-time forward at inference you must tell the model which positions are prefix.

In the current Transformers port the mask is controlled by token_type_ids:

• token_type_ids[i] == 1 → position i is part of the prefix block (bidirectional within the block).
• otherwise → causal.

If you omit token_type_ids, attention falls back to pure causal, which does not match the pre-training distribution and will give noticeably worse logits. The simplest correct call passes token_type_ids = torch.ones_like(input_ids), marking the entire input prompt as one bidirectional prefix block — exactly how training-time prefill ran.

Architecture

The recurrent core (per forward pass, in inference mode):

z_H = embed(input_ids) * embedding_scale
z_L = z_L_init.expand_as(z_H)

for _ in range(H_cycles):
    for _ in range(L_cycles):
        z_L = L_module(z_L + z_H)
    z_H = H_module(z_H + z_L)
return z_H

Both stacks share the same Transformer block design (gated attention, RoPE, SwiGLU, pre-RMSNorm); see Model details above for shapes.

Training data

Pre-trained on a sampled mixture of publicly available text corpora. The full dataset composition, sampling weights, and preprocessing pipeline are open-sourced:

Limitations

• English only (training corpus is predominantly English).
• HRM-Text-1B was not trained on code datasets, therefore its rather weak performance on coding tasks was expected. Early third-party code SFT experiments on roughly 1B tokens of code data improved coding benchmark scores from low single digits to around 40–50, suggesting promising adaptation potential, but those results are not part of this checkpoint.
• Outputs may vary under differen