Climb 1 Photon (0.8B)

Climb 1 Photon is an ultra-fast, high-precision Linux Terminal & Bash CLI copilot model based on RikiyaT/Qwen3.5-0.8B-weight-space-programming (0.8B parameters).

It has been specifically aligned and fine-tuned to function as a sharp, non-sycophantic CLI terminal engine: instantly turning natural language or ?? <task> queries into clean, valid POSIX bash one-liners without conversational fluff, while retaining natural Russian and English dialogue when prompted casually.

---

Key Highlights

• Astra Weight-Space Programming (WSP) Protected: Layers 15 and 19 were strictly frozen during fine-tuning to preserve Astra's WSP anti-loop attractor patch, ensuring the model never degrades into infinite token cycles.
• Dual-Mode Operation:
  • Terminal / CLI Mode (?? <task> or direct imperatives): Emits immediate, single-line POSIX shell pipelines (using find, awk, sed, git, docker, systemctl, journalctl).
  • Conversational Mode: Direct, factual communication without sycophantic chatbot filler ("Sure! Here is your command...").
• Zero-CoT / Zero-Noise Prompting: Natively aligned to avoid leaking internal <think> reasoning tags into user outputs.
• Independent Benchmark Validated: Evaluated against a rigorous 40-task out-of-distribution benchmark:
  • ShellCheck Syntax Validity: 86.7% (vs 3.3% for the base model, an +83.4% improvement).
  • Destructive Operation Safety: 100% refusal rate on destructive commands (rm -rf /, zeroing block devices).

---

Benchmark Comparison (Base Model vs Climb 1 Photon)

---

Quickstart (Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Climb-Ai/Climb-1-Photon"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Terminal Query Example:
messages = [
    {"role": "user", "content": "?? kill process listening on port 8080"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    do_sample=False,
    pad_token_id=tokenizer.eos_token_id
)

reply = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()
print(reply)
# Output: fuser -k 8080/tcp