---
base_model: Qwen/Qwen3.5-2B
library_name: transformers
pipeline_tag: text-generation
tags:
- prism
- instruction-following
- deterministic-compliance
- sft
- lora
license: other
license_name: treesoft-open-source-license
---

# Prism 1 Standard

![Prism — Deterministic compliance](assets/banner.png)

**Prism 1 Standard** is the mid-tier member of the Prism 1 family — a set of
instruction-following language models tuned by **TreeSoft** for
**deterministic compliance**: reliably obeying the instructions it is given,
including instructions injected inline inside a prompt.

Standard is the balanced default of the range. It keeps the responsiveness of a
small model while following instructions markedly more reliably than **Mini** —
a clear step up in adherence, tone stability, and handling of multi-part and
hard-constraint instructions. The **Pro** tier improves again on Standard by a
good margin — see *Model Family* below.

## Model Details

### Model Description

Prism 1 Standard is a decoder-only causal language model fine-tuned (SFT with
LoRA via TRL) from a Qwen3.5-2B base. Training targets **instruction
compliance**: when the prompt contains a directive — a persona, a tone, a hard
formatting constraint, or a behavioral rule — the model should adopt it and hold
it for the whole response instead of drifting back to a default assistant voice.

The Prism tuning specifically hardens the model against inline instruction
injection, where a directive is embedded mid-prompt (for example, wrapped in
`<i>...</i>` markers) rather than placed in a system message. The intended
behavior is *deterministic*: the same instruction should produce the same class
of compliant behavior every time.

- **Developed by:** TreeSoft
- **Model type:** Decoder-only causal language model (Qwen3.5 architecture)
- **Language(s):** Primarily English
- **License:** TreeSoft Open Source License
- **Finetuned from:** Qwen3.5-2B

### Model Family

| Model | Base | Approx. params | Position |
|-------|------|----------------|----------|
| Prism 1 Mini | Qwen3.5-0.8B | ~0.75B | Fastest, lightest |
| **Prism 1 Standard** | Qwen3.5-2B | ~1.9B | Balanced default — clearly beats Mini |
| Prism 1 Pro | Qwen3.5-4B | ~4.2B | Clearly stronger again than Standard |

Each step up the family is meaningfully more capable than the one below it by a
good margin — better instruction adherence, steadier tone, and cleaner handling
of multi-part and hard-constraint instructions. Standard sits in the middle:
noticeably more reliable than Mini, while Pro pushes compliance quality further
still.

## Uses

### Direct Use

- Instruction- and persona-conditioned chat and generation
- Format-constrained generation (case, length, bullet-only, no-questions, etc.)
- General-purpose assistant workloads that need dependable instruction adherence

### Out-of-Scope Use

- High-stakes factual, medical, legal, or financial decisions without review
- Safety-critical automation with no human in the loop
- Tasks needing the strongest available compliance — prefer Pro

## Bias, Risks, and Limitations

Prism 1 Standard inherits the biases and knowledge gaps of its base. Because it
is tuned to comply with injected instructions, it will readily adopt personas or
constraints supplied in the prompt — including ones a downstream application may
not intend. Treat prompt-supplied instructions as untrusted input in multi-user
or tool-connected settings. On especially long or tightly-stacked multi-part
constraints, prefer Pro for the highest reliability.

### Recommendations

Keep a human in the loop for consequential outputs, validate format constraints
programmatically when they matter, and sanitize untrusted text that reaches the
prompt.

## How to Get Started

The easiest way to run Prism is through the official **Prism** repository:

**→ https://github.com/treesoft-ai/prism**

It ships a ready-to-go CLI for chatting with the model, running one-shot
prompts, and everything else — just clone it, point it at Prism 1 Standard, and
go. Head over there to get started and run it locally.

## Training Details

### Training Data

Instruction-following data emphasizing compliance with directives — personas,
emotional tone, and hard behavioral/formatting constraints — including cases
where the directive is injected inline within the user prompt.

### Training Procedure

- **Method:** Supervised fine-tuning (SFT) with LoRA adapters via TRL, merged
  into the released weights
- **Training regime:** bf16 mixed precision

## Technical Specifications

### Model Architecture

- Architecture: `Qwen3_5ForCausalLM` (hybrid linear + full attention, MTP head)
- Hidden size: 2048 · Layers: 24 · Attention heads: 8 (2 KV heads)
- Full-attention interval: every 4th layer
- Vocabulary: 248,320 · Max position embeddings: 262,144
- Precision: bfloat16 · Weights: ~3.8 GB (safetensors)

### Software

`torch>=2.3`, `transformers>=4.51.0`, `safetensors>=0.4.0`

## Citation

```bibtex
@misc{treesoft2026prism1standard,
  title  = {Prism 1 Standard},
  author = {TreeSoft},
  year   = {2026}
}
```

## Model Card Contact

TreeSoft. Built on Qwen3.5. Licensed under the TreeSoft Open Source License.