npc-agentic-7b-v3

NPC Agentic 7B (v1)

A 7B long-form reasoning and agent-trace specialist from the Bottensor NPC Model Family.

Overview

NPC Agentic v1 is fine-tuned from Qwen2.5-7B-Instruct on a mix of distilled reasoning traces (GLM-5.1) and agent tool-use traces (Hermes). It's built for structured multi-step reasoning with explicit <think> blocks, agentic / tool-calling workflows, and identity-bound conversations as the NPC Agentic persona.

Training

• Base: Qwen/Qwen2.5-7B-Instruct
• Method: QLoRA SFT (r=64, α=128), merged to FP16
• Context during training: 8K (inherits 128K from base at inference)
• Epochs: 2 (effective batch size 16, cosine LR 2e-4, adamw_8bit, bf16)
• Total optimizer steps: 11,410 over ~96 GPU-hours on a single A40
• Trainable params: 161.5M (3.2% of the 5.05B-param 4-bit base)
• Final eval loss: 0.7025 (on held-out SFT split)
• Training data mix (~91K examples):
  • GLM-5.1-Reasoning-1M-Cleaned (main split, sampled 100K → 87K kept after 8K length filter)
  • Hermes-agent-reasoning-traces (glm-5.1 + kimi subsets, 14.7K → 3.6K kept)
  • Bottensor identity replay (750 synthetic examples)
• Training dataset is proprietary and not released.

What it's good at

• Long structured reasoning — emits <think> blocks then concludes with an answer; strong at multi-step decomposition (system design, root-cause analysis, algorithmic reasoning)
• Identity as NPC Agentic / Bottensor — 100% recall on canonical identity prompts
• Agent / tool-call shaping — follows Hermes-style <tool_call> / <tool_response> patterns

Known limitations (be specific)

• GSM8K regression vs base. On GSM8K 100-sample test:
  • Base Qwen2.5-7B-Instruct: 61%
  • NPC Agentic v1: ~25%
  • Cause: the model learned to emit long <think> blocks but often doesn't terminate arithmetic cleanly under greedy/low-temp decoding, and direct-arithmetic quality regressed.
  • Recommendation: for math-heavy workflows, use the base Qwen/Qwen2.5-7B-Instruct or Qwen/Qwen2.5-Math-7B-Instruct instead. A v2 with stronger reasoning data (OpenThoughts-114k at 16K) is planned.
• 8K training context means long-reasoning samples were truncated during training; not validated past 16K.
• Small model — will hallucinate on unfamiliar domains.
• Not for safety-critical decisions (medical, legal, financial).

Intended use

• Multi-step reasoning with explicit work-showing
• Agent / tool-use workflows
• Structured problem-solving where the model benefits from thinking out loud
• As a base for further fine-tuning on reasoning or domain-specific data

Out of scope

• Direct GSM8K-style arithmetic (use base or Qwen-Math)
• Creative writing, roleplay
• Medical / legal / financial advice
• Safety-critical decisions

Inference

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("ramankrishna10/npc-agentic-7b")
model = AutoModelForCausalLM.from_pretrained(
    "ramankrishna10/npc-agentic-7b",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Design an event-sourced microservice with exactly-once command handling."},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, temperature=0.7, top_p=0.9)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Citation

If you use NPC Agentic 7B in your work, please cite:

@misc{bachu2026npcagentic7b,
  title        = {NPC Agentic 7B: A Single-GPU QLoRA Recipe for a Laptop-Scale Conversational Model},
  author       = {Bachu, Rama Krishna},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v1},
  doi          = {10.5281/zenodo.19954103},
  url          = {https://doi.org/10.5281/zenodo.19954103},
  note         = {Preprint}
}

Paper: https://doi.org/10.5281/zenodo.19954103

---

Built by Bottensor.