Eval: HRM-Text-1B architecture for email classification fine-tuning #70

New issue

Open

opened 2026-07-09 23:05:32 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-07-09 23:05:32 -07:00

Owner

Source: https://huggingface.co/sapientinc/HRM-Text-1B
License: Apache 2.0
Size: 1B params, novel Hierarchical Reasoning Model (HRM) architecture

HRM uses dual-timescale recurrent processing (two iterating Transformer modules). Strong at classification, extraction, and structured output with few-shot examples. Pre-alignment checkpoint — needs fine-tuning for production use.

CF relevance:

Avocet's email classifier could benefit from a model optimized for classification/extraction
1B params is lean enough to run on Heimdall or Sif for fine-tuning experiments
Apache 2.0 license — clean for Avocet's BSL fine-tuned weights model

Evaluate:

Benchmark on Avocet's existing labeled email dataset vs. current base model
HRM's structured output strength may help with multi-label classification (intent + urgency + category)
English-only, weak coding — acceptable for email classification use case
Pre-alignment means it will need fine-tuning; compare fine-tune cost vs. current approach

Priority: Low — evaluate after current benchmark harness stabilizes.

**Source:** https://huggingface.co/sapientinc/HRM-Text-1B **License:** Apache 2.0 **Size:** 1B params, novel Hierarchical Reasoning Model (HRM) architecture HRM uses dual-timescale recurrent processing (two iterating Transformer modules). Strong at classification, extraction, and structured output with few-shot examples. Pre-alignment checkpoint — needs fine-tuning for production use. **CF relevance:** - Avocet's email classifier could benefit from a model optimized for classification/extraction - 1B params is lean enough to run on Heimdall or Sif for fine-tuning experiments - Apache 2.0 license — clean for Avocet's BSL fine-tuned weights model **Evaluate:** - Benchmark on Avocet's existing labeled email dataset vs. current base model - HRM's structured output strength may help with multi-label classification (intent + urgency + category) - English-only, weak coding — acceptable for email classification use case - Pre-alignment means it will need fine-tuning; compare fine-tune cost vs. current approach **Priority:** Low — evaluate after current benchmark harness stabilizes.