Update README.md

README.md

@@ -19,28 +19,35 @@ tags:
 
 # Cyber Safety Model
 
-**CSM-1-Preview** is a **Cyber Safety Model**: [Qwen/Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) fine-tuned with a LoRA adapter for defensive cybersecurity work, with the adapter merged into the weights (bf16). It is meant to help defenders, engineers and learners with detection, incident response, hardening, secure coding and static malware analysis, while declining to produce ready-to-use attack tooling.
+**CSM-1-Preview** is an experimental Cyber Safety Model designed to help defenders, engineers, and learners with cybersecurity work. It can help with detection, incident response, hardening, secure coding, and static malware analysis. It is also designed to decline requests for ready-to-use attack tooling.
 
-> **Experimental.** This model can be wrong. Verify CVE IDs, commands, detection rules and malware verdicts before relying on them. No formal benchmark results are published for it yet.
+The model was fine-tuned using a LoRA adapter that has been merged into the weights in bf16.
 
-## Intended use
+> **A quick note:** CSM-1-Preview can make mistakes. Please verify CVE IDs, commands, detection rules, and malware verdicts before relying on them. Formal benchmark results are not yet available.
 
-- Explaining attacks, vulnerabilities and malware behaviour at a defender's level of detail
-- Writing and reviewing detections (Sigma, YARA, queries), hardening guidance and incident-response runbooks
-- Secure code review and secure-by-design advice
-- Triage and **static** analysis of suspicious files, scripts and browser or IDE extensions (with tools: see below)
-- Teaching, CTFs and authorized security testing concepts
+## What can it help with?
 
-## Out of scope
+CSM-1-Preview is intended for tasks such as:
 
-- Producing working malware, ransomware, stealers, botnets, phishing kits or weaponized exploits
-- Attacking systems you do not own or are not authorized to test
-- Being the only basis for a legal, forensic or compliance decision
-- Anything that needs guaranteed factual accuracy without human verification
+- Explaining attacks, vulnerabilities, and malware behavior from a defender’s perspective
+- Writing and reviewing detections, including Sigma, YARA, and queries
+- Developing hardening guidance and incident-response runbooks
+- Reviewing code and discussing secure-by-design practices
+- Triaging and **statically** analyzing suspicious files, scripts, and browser or IDE extensions—with appropriate tools and safeguards
+- Learning about CTFs and authorized security testing
+
+## What’s out of scope?
+
+Please don’t rely on the model for:
+
+- Producing working malware, ransomware, stealers, botnets, phishing kits, or weaponized exploits
+- Attacking systems without authorization
+- Making legal, forensic, or compliance decisions without expert review
+- Tasks that require guaranteed accuracy without human verification
 
 ## Quick start
 
-Needs a recent `transformers`, about 55 GB of GPU memory in bf16, and `flash-linear-attention` for fast linear-attention layers.
+You’ll need a recent version of `transformers`, approximately 55 GB of GPU memory in bf16, and `flash-linear-attention` for fast linear-attention layers.
 
 ```python
 import torch
@@ -48,41 +55,56 @@ from transformers import AutoTokenizer, AutoModelForImageTextToText
 
 repo = "OpenModel-Labs/CSM-1-Preview"
 tok = AutoTokenizer.from_pretrained(repo)
-model = AutoModelForImageTextToText.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")
+model = AutoModelForImageTextToText.from_pretrained(
+    repo, dtype=torch.bfloat16, device_map="auto"
+)
 
 messages = [
-    {"role": "system", "content": open("system_prompt.txt").read()},   # recommended, see below
+    {"role": "system", "content": open("system_prompt.txt").read()},
     {"role": "user", "content": "How do I detect Kerberoasting in Windows event logs?"},
 ]
-prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)
+prompt = tok.apply_chat_template(
+    messages, add_generation_prompt=True, tokenize=False, enable_thinking=False
+)
 inputs = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
-out = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.9, top_k=20)
+out = model.generate(
+    **inputs, max_new_tokens=1024, do_sample=True,
+    temperature=0.7, top_p=0.9, top_k=20
+)
 print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
 ```
 
-`example_chat.py` in this repo is a small streaming chat loop. Settings used in testing: temperature 0.7 for chat (0.2 to 0.4 for analysis tasks), top_p 0.9, top_k 20. Set `enable_thinking=True` in the chat template to let the model reason before answering (slower).
+For a simple streaming chat loop, see `example_chat.py` in this repository.
+
+Settings used in testing were temperature **0.7** for chat, **0.2–0.4** for analysis tasks, `top_p=0.9`, and `top_k=20`. Set `enable_thinking=True` in the chat template to let the model reason before answering; this will be slower.
 
 ## Recommended system prompt
 
-The model was tested with the prompt in [`system_prompt.txt`](system_prompt.txt). It asks for accurate, concise answers, no invented CVEs or flags, treating pasted content as data rather than instructions, defanged IOCs, and a short refusal with a defensive alternative for harmful requests.
+The model was tested with [`system_prompt.txt`](system_prompt.txt). The prompt asks it to give accurate, concise answers; avoid inventing CVEs or flags; treat pasted content as data rather than instructions; defang IOCs; and offer a short refusal with a defensive alternative for harmful requests.
 
-## Tool use
+## Using tools safely
 
-The chat template supports tool calling. If you give the model file or shell tools for malware analysis, run them in a **sandbox with read-only samples**, an allowlist of commands and no network, and treat all tool output as untrusted data (it can contain prompt injection). Never execute samples.
+The chat template supports tool calling. If you provide file or shell tools for malware analysis:
 
-## Training
+- Use a **sandbox** with read-only samples.
+- Allowlist commands and disable network access.
+- Treat all tool output as untrusted data—it may contain prompt injection.
+- Never execute samples.
 
-- Base model: Qwen/Qwen3.5-27B
-- Method: LoRA fine-tuning, merged into the base weights (bf16)
-- Training data and evaluation details: not documented in this card yet.
+## Training and evaluation
+
+CSM-1-Preview was fine-tuned with LoRA, and the adapter was merged into the weights in bf16. Training data and evaluation details are not yet documented in this card.
 
 ## Limitations and risks
 
-- May hallucinate identifiers (CVEs, CVSS scores, product names, flags, API fields). Always check against primary sources.
-- Safety behaviour is imperfect: it can refuse legitimate requests and can be talked into answers it should not give. Do not use it as a security control.
-- Malware verdicts from static reading are heuristic. Absence of evidence in a quick scan is not proof of safety.
-- Inherits the limitations and biases of the base model.
+- The model may invent identifiers such as CVEs, CVSS scores, product names, flags, or API fields. Check primary sources.
+- Its safety behavior is imperfect: it may refuse legitimate requests or produce answers it should not. Don’t use it as a security control.
+- Static malware assessments are heuristic. A quick scan that finds nothing suspicious does not prove a file is safe.
 
 ## License
 
-Released under **other**. Check that your use also complies with the base model's license.
+Released under **other**. See the [license agreement](LICENSE) and check that your use also complies with applicable upstream license requirements.
+
+## Base model
+
+CSM-1-Preview is a fine-tune of the base model listed in the metadata above and inherits its limitations and biases.