Instructions to use sanjayram-a/nanoagent-15m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sanjayram-a/nanoagent-15m with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="sanjayram-a/nanoagent-15m")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("sanjayram-a/nanoagent-15m", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use sanjayram-a/nanoagent-15m with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sanjayram-a/nanoagent-15m"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sanjayram-a/nanoagent-15m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sanjayram-a/nanoagent-15m

SGLang

How to use sanjayram-a/nanoagent-15m with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "sanjayram-a/nanoagent-15m" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sanjayram-a/nanoagent-15m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "sanjayram-a/nanoagent-15m" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sanjayram-a/nanoagent-15m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sanjayram-a/nanoagent-15m with Docker Model Runner:
```
docker model run hf.co/sanjayram-a/nanoagent-15m
```

NanoAgent-15M

A 15M-parameter Llama-style decoder-only transformer for chat and structured tool calling, trained end-to-end on a single RTX 3050 Laptop GPU (4GB VRAM). Built for local edge deployment (~58MB per fp32 checkpoint).

Full training code + data pipeline: https://github.com/sanjayram-a/nanoagent-15m
Interactive training curves for every run below live under this repo's Training metrics tab

This is a 15M model. Grade the format (stops at <|im_end|>, well-formed <tool_call>{…}</tool_call> blocks, stays on topic a sentence or two), not GPT-level polish.

Model details


Architecture	Llama-style decoder-only (`transformers.LlamaForCausalLM`)
Hidden / layers / heads	384 / 8 / 6 Q + 2 KV (GQA 3:1), SwiGLU 1024
Context length	1024 tokens (prompt + reply combined)
Vocab	6,144 byte-level BPE, custom-trained (`nanoagent_tokenizer/`)
Positional encoding	RoPE (theta 10,000); RMSNorm; tied word embeddings
Parameters	~~15M (~~58MB fp32 `.pt`)
Chat format	`<\|im_start\|>{role}\n{content}<\|im_end\|>`; tools in `<tools>…</tools>`; calls in `<tool_call>{…}</tool_call>`; results in `<tool_response>…</tool_response>`

Checkpoints in this repo

File	Variant	Val loss	Val ppl
`runs/nanoagent_15m/final_15m_model.pt`	Base (pretrained, no SFT — use with `chat.py --raw`)	3.9772	53.37
`runs/nanoagent_15m/best_model.pt`	Base, best-val snapshot	3.9791	~53.5
`runs/nanoagent-15m-sft/sft_chat_max_final.pt`	SFT run 1 (blended 409,760-row mix) — recommended chat + tool weights	1.7348	5.67
`runs/nanoagent-15m-sft2/checkpoints/latest.pt`	SFT run 2 / instruct, full training checkpoint (`model_state` + optimizer)	1.4783	4.39
`nanoagent_tokenizer/`	Tokenizer (`tokenizer.json`, `tokenizer_config.json`) — required by every checkpoint	—	—

Base pretraining — `final_15m_model.pt`

Trained tokens (read from TensorBoard + `training_summary.json`)

Metric	Value
Tokens seen	999,948,288 (~1B)
Optimizer steps	15,256 (batch 16 × 1024 × grad-accum 4 = 65,536 tok/step)
Stage 1 → Stage 2	3,051 steps @ 200M tokens → remainder @ 800M tokens
Train loss	8.18 → 4.02
Val loss	6.19 → 3.98 (best 3.9791, final 3.9772, ppl 53.37)
Throughput	~37–38K tok/s on RTX 3050 Laptop 4GB
LR schedule	warmup 1,000 steps to 6e-4, cosine decay to 6e-5
Interruption	crashed once at step 8,239 (`inf` grad norm, fp16 overflow), resumed from emergency checkpoint to completion

Datasets (from `curriculum_meta.json` + run log)

Packed to train.bin (1,000,000,000 tokens, 2000MB) + val.bin (10,000,000 tokens, 20MB). Acquisition targets below include +15% selection and +15% acquisition margins; the sampler never shuffles across the 200M-token stage boundary.

Source	HF dataset (config / split)	Train target (tokens)	Val target (tokens)	Stage 1 share	Stage 2 share
l3	`openbmb/Ultra-FineWeb-L3` (Ultra-FineWeb-L3-en-Multi-Style-Synthetic / train)	409,975,000	3,967,500	35%	30%
fineweb	`HuggingFaceFW/fineweb-edu` (sample-10BT / train, score ≥ 3)	357,075,000	3,306,250	35%	25%
cosmo	`HuggingFaceTB/smollm-corpus` (cosmopedia-v2 / train)	330,625,000	3,306,250	25%	25%
code	`openbmb/UltraData-Code` (UltraData-Code-L2 / py)	224,825,000	2,645,000	5%	20%

Documents selected: 12,102 validation / 251,506 stage-1 / 969,804 stage-2.

SFT run 1 — `sft_chat_max_final.pt` (recommended)

Trained tokens (read from TensorBoard)

Metric	Value
Tokens seen	409,544,800 (~410M, ~1 epoch over the mix)
Optimizer steps	12,500
Train loss	6.62 → 1.70
Val loss	2.36 → 1.79 (final 1.7348, ppl 5.67)

Datasets (409,760 raw rows, single blended pass + 15% tool-context negatives)

Source	HF dataset (config)	Rows	Share
instruct	`HuggingFaceTB/smoltalk` (smol-magpie-ultra)	150,000	36.6%
worldqa	`HuggingFaceTB/ultrachat_questions_about_world`	86,000	21.0%
xlam	`Salesforce/xlam-function-calling-60k` (gated — needs an accepted-terms token)	60,000	14.6%
glaive	`glaiveai/glaive-function-calling-v2`	47,500	11.6%
summarize	`HuggingFaceTB/smoltalk` (smol-summarize)	34,000	8.3%
rewrite	`HuggingFaceTB/smoltalk` (smol-rewrite)	30,000	7.3%
chat	`HuggingFaceTB/smoltalk` (everyday-conversations)	2,260	0.6%

Tool-negative injection: 15% of non-tool rows keep their plain-text answer but get an unrelated tool list attached, teaching that tools in context don't always need calling.

SFT run 2 / instruct — `checkpoints/latest.pt`

Trained tokens (read from TensorBoard)

Metric	Value
Tokens seen	895,970,816 (~896M)
Optimizer steps	27,350
Train loss	6.63 → 1.53
Val loss	2.19 → 1.48 (latest 1.4783, ppl 4.39)
LR schedule	warmup 200 steps to 5e-5, cosine decay to 5e-6, grad-accum 4, 8192-token batch budget

Datasets (on-disk tokenized cache `./sft_cache`: 380,573 train / 1,939 val rows, 3 epochs)

Source	HF dataset	Row cap
instruct	`HuggingFaceTB/smoltalk` (smol-magpie-ultra)	200,000
glaive	`glaiveai/glaive-function-calling-v2`	90,000
xlam	`Salesforce/xlam-function-calling-60k` (gated)	60,000
summarize	`HuggingFaceTB/smoltalk` (smol-summarize)	40,000
rewrite	`HuggingFaceTB/smoltalk` (smol-rewrite)	40,000
chat	`HuggingFaceTB/smoltalk` (everyday-conversations)	2,260
worldqa	— (excluded from this run)	0

Tool-negative ratio 0.10. Lowest validation numbers of all three runs.

Usage

Chat script + tool definitions live in the GitHub repo (chat.py, tools.json). Checkpoints here load with the same architecture:

# Chat + tools (SFT weights)
python chat.py --model_path runs/nanoagent-15m-sft/sft_chat_max_final.pt --tools tools.json

# Single-shot test
python chat.py --model_path runs/nanoagent-15m-sft/sft_chat_max_final.pt \
  --tools tools.json --prompt "What is the weather in Paris?" \
  --max_new_tokens 40 --temperature 0

# Base weights (never saw the chat template — use --raw)
python chat.py --model_path runs/nanoagent_15m/final_15m_model.pt --raw --prompt "Hello world"

Minimal transformers loading (matches train.py / chat.py exactly)

# Minimal transformers loading (matches train.py / chat.py exactly)
from transformers import LlamaConfig, LlamaForCausalLM, PreTrainedTokenizerFast
import torch

config = LlamaConfig(vocab_size=6144, hidden_size=384, intermediate_size=1024,
                     num_hidden_layers=8, num_attention_heads=6, num_key_value_heads=2,
                     max_position_embeddings=1024, rms_norm_eps=1e-5,
                     rope_theta=10000.0, tie_word_embeddings=True)
model = LlamaForCausalLM(config)
state = torch.load("runs/nanoagent-15m-sft/sft_chat_max_final.pt", map_location="cpu")
model.load_state_dict(state["model_state"] if "model_state" in state else state)
tokenizer = PreTrainedTokenizerFast.from_pretrained("nanoagent_tokenizer")

Limitations

15M parameters / 1024-token context: short, simple exchanges only; weak factuality.
Tool-call format is the skill being taught — always validate emitted JSON before executing.
xlam-derived rows are ~9% malformed upstream (skipped + counted during SFT).
No reasoning/scratchpad tokens by design at this size; no QK-Norm (stock Llama has none).

Reproducing / full logs

The GitHub repo contains the full pipeline (train-tokenizer.py, download_and_pack.py, train.py, sft_chat.py, sft_common.py, sft_agent_openbmb.py, sft-chat-max.py), the raw run notebook (note.md), per-run TensorBoard events (runs/*/events.out.tfevents.*, mirrored in the Training metrics tab), and training_config.json / training_summary.json for the base run.

Downloads last month: -; Downloads are not tracked for this model. How to track

Evaluation results

final validation loss on NanoAgent 1B-token pretraining mix (Ultra-FineWeb-L3 + FineWeb-Edu + Cosmopedia-v2 + UltraData-Code)
self-reported

3.977
final validation perplexity on NanoAgent 1B-token pretraining mix (Ultra-FineWeb-L3 + FineWeb-Edu + Cosmopedia-v2 + UltraData-Code)
self-reported

53.370
final validation loss on Blended SFT mix, 409,760 rows (smoltalk + ultrachat world-QA + xlam-function-calling-60k + glaive-function-calling-v2 + summarize + rewrite)
self-reported

1.735
final validation perplexity on Blended SFT mix, 409,760 rows (smoltalk + ultrachat world-QA + xlam-function-calling-60k + glaive-function-calling-v2 + summarize + rewrite)
self-reported

5.670
latest validation loss on Instruct-heavy SFT mix, 380,573 tokenized train rows (smol-magpie-ultra + glaive + xlam + summarize + rewrite + chat)
self-reported

1.478
latest validation perplexity on Instruct-heavy SFT mix, 380,573 tokenized train rows (smol-magpie-ultra + glaive + xlam + summarize + rewrite + chat)
self-reported

4.390

NanoAgent-15M

Model details

Checkpoints in this repo

Base pretraining — final_15m_model.pt

Trained tokens (read from TensorBoard + training_summary.json)

Datasets (from curriculum_meta.json + run log)

SFT run 1 — sft_chat_max_final.pt (recommended)

Trained tokens (read from TensorBoard)

Datasets (409,760 raw rows, single blended pass + 15% tool-context negatives)

SFT run 2 / instruct — checkpoints/latest.pt

Trained tokens (read from TensorBoard)

Datasets (on-disk tokenized cache ./sft_cache: 380,573 train / 1,939 val rows, 3 epochs)

Usage

Limitations

Reproducing / full logs

Evaluation results

Base pretraining — `final_15m_model.pt`

Trained tokens (read from TensorBoard + `training_summary.json`)

Datasets (from `curriculum_meta.json` + run log)

SFT run 1 — `sft_chat_max_final.pt` (recommended)

SFT run 2 / instruct — `checkpoints/latest.pt`

Datasets (on-disk tokenized cache `./sft_cache`: 380,573 train / 1,939 val rows, 3 epochs)