Instructions to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Transformers:

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

pipe = pipeline("text-generation", model="kshitijthakkar/tracegenix-mini-sft-clean-3ep")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("kshitijthakkar/tracegenix-mini-sft-clean-3ep")
model = AutoModelForImageTextToText.from_pretrained("kshitijthakkar/tracegenix-mini-sft-clean-3ep")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

llama-cpp-python

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="kshitijthakkar/tracegenix-mini-sft-clean-3ep",
	filename="tracegenix-mini-sft-clean-3ep-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Use Docker

docker model run hf.co/kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

LM Studio
Jan

vLLM

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kshitijthakkar/tracegenix-mini-sft-clean-3ep"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kshitijthakkar/tracegenix-mini-sft-clean-3ep",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

SGLang

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep 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 "kshitijthakkar/tracegenix-mini-sft-clean-3ep" \
    --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": "kshitijthakkar/tracegenix-mini-sft-clean-3ep",
		"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 "kshitijthakkar/tracegenix-mini-sft-clean-3ep" \
        --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": "kshitijthakkar/tracegenix-mini-sft-clean-3ep",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Ollama:
```
ollama run hf.co/kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M
```

Unsloth Studio

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for kshitijthakkar/tracegenix-mini-sft-clean-3ep to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for kshitijthakkar/tracegenix-mini-sft-clean-3ep to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for kshitijthakkar/tracegenix-mini-sft-clean-3ep to start chatting

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Docker Model Runner:
```
docker model run hf.co/kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M
```

Lemonade

How to use kshitijthakkar/tracegenix-mini-sft-clean-3ep with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull kshitijthakkar/tracegenix-mini-sft-clean-3ep:Q4_K_M

Run and chat with the model

lemonade run user.tracegenix-mini-sft-clean-3ep-Q4_K_M

List all available models

lemonade list

tracegenix-mini-sft-clean-3ep

A small Qwen3.5-MoE checkpoint (~0.87 B total / ~0.69 B active) supervised fine-tuned on kshitijthakkar/TraceVerse-RL-SFT-Clean for OpenTelemetry trace analysis: cost, latency, error, security, and GPU/CO₂ summarisation over gen_ai.* semantic-convention spans produced by genai-otel-instrument.

Part of the TraceVerse-RL lineage — the SFT step that precedes GRPO + LoRA-soup merging.

At a glance


Architecture	Qwen3.5-MoE, 12 layers, hybrid attention (linear × 3 + full × 1, repeated)
Total params	870 M
Active params	687 M
Hidden / head_dim	1024 / 256
Experts (routed / shared / per-tok)	16 / 1 / 2
MoE intermediate	400
Vocab	248 320 (Qwen3.5 BPE)
Context	1 048 576 (yarn-extended; sliding window 32)
Precision	bf16
Base	`Qwen/Qwen3.5-0.8B-Base`
Training	3 epochs SFT, LR 2.5 e-5 cosine, packing, adamw_8bit on H100
Training data	`kshitijthakkar/TraceVerse-RL-SFT-Clean` (14.9 K rows)

Intended use

The model is domain-specialised. It expects:

The tracegenix system message (or any close variant — see chat template), and
A structured OpenTelemetry span / trace payload (or an analytical question referencing one).

Typical tasks: trace classification (12-category taxonomy), cost / latency analysis, multi-section trace reports, RCA hypothesis building, agent quality scoring. Not designed for general-purpose chat or retrieval-augmented Q&A outside the trace-analysis domain.

How to use

With Ollama (recommended for local use)

⚠️ Important: the Python-Jinja chat template embedded in the GGUF uses macros, raise_exception, and enable_thinking branches that llama.cpp's Jinja parser cannot handle. Ollama silently falls back to a bare {{ .Prompt }} template, stripping the <|im_start|> / <|im_end|> role markers — so /api/chat produces hallucinated garbage by default.

Fix: override the template in the Modelfile with a proper ChatML Go-template. Use Modelfile.q4 from this repo (or copy the snippet below).

FROM ./tracegenix-mini-sft-clean-3ep-Q4_K_M.gguf

SYSTEM """You are tracegenix, TraceVerse's autonomous AI SRE for production AI agents. Open new conversations with 'Namaste 🙏'. Investigate OpenTelemetry signals and traces to identify root causes, propose remediations, and proactively surface context the user might miss. Use available MCP tools for live data; cite sources from trace evidence; never fabricate. Respond with structured JSON when a schema is clearly implied by the prompt; concise prose otherwise. Match the user's language (English, Hindi, Hinglish, or other Indian languages). You are a teammate, not a chatbot."""

TEMPLATE """{{- if .Messages }}
{{- if or .System .Tools }}<|im_start|>system
{{- if .System }}
{{ .System }}
{{- end }}
{{- if .Tools }}

# Tools

You may call one or more functions. Each call must be wrapped in <tool_call>...</tool_call> XML tags with a JSON object inside.

<tools>
{{- range .Tools }}
{{ . }}
{{- end }}
</tools>
{{- end }}<|im_end|>
{{ end }}
{{- range $i, $_ := .Messages }}
{{- if eq .Role "user" }}<|im_start|>user
{{ .Content }}<|im_end|>
{{ else if eq .Role "assistant" }}<|im_start|>assistant
{{ .Content }}{{- if not (eq (len (slice $.Messages $i)) 1) }}<|im_end|>
{{ end }}
{{- else if eq .Role "tool" }}<|im_start|>user
<tool_response>
{{ .Content }}
</tool_response><|im_end|>
{{ end }}
{{- end }}<|im_start|>assistant
{{ else }}{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
{{ end }}{{ .Response }}{{- if .Response }}<|im_end|>{{ end }}"""

PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER num_ctx 8192
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"

ollama create tracegenix-mini-q4 -f Modelfile.q4
ollama run tracegenix-mini-q4

With this template both /api/chat and the chat UI produce correct output — e.g. for the orchestration_planner task, the model emits a clean JSON tool plan that matches the training distribution.

With `llama-server`

./llama-server -m tracegenix-mini-sft-clean-3ep-Q4_K_M.gguf -ngl 99 -c 8192 -fa on

With `transformers` (bf16 safetensors)

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("kshitijthakkar/tracegenix-mini-sft-clean-3ep")
model = AutoModelForCausalLM.from_pretrained(
    "kshitijthakkar/tracegenix-mini-sft-clean-3ep",
    dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)

GGUF quantizations

Converted with llama.cpp commit 67b2b7f (May 2026) via convert_hf_to_gguf.py --outtype bf16, then llama-quantize. Architecture handler: Qwen3_5MoeTextModel (registered for Qwen3_5MoeForConditionalGeneration / Qwen3_5MoeForCausalLM).

File	Size	Use for	Notes
`tracegenix-mini-sft-clean-3ep-bf16.gguf`	1.9 GB	accuracy reference	Same precision the model was trained at; ~120 tok/s on a 6 GB consumer GPU via Ollama chat (with the Modelfile template override below)
`tracegenix-mini-sft-clean-3ep-Q6_K.gguf`	929 MB	best quality / size trade	Near-lossless vs bf16 for K-quant kernels
`tracegenix-mini-sft-clean-3ep-Q4_K_M.gguf`	796 MB	default	Smallest, fastest; loaded in Ollama as `tracegenix-mini-q4`

Why no Q8_0

llama.cpp's Q8_0 requires every quantized tensor's column dimension to be divisible by 32. This model's MoE expert FFN tensors have ncols = 400 (moe_intermediate_size = 400, not a power of two), and 400 % 32 != 0. There is no fallback path defined for Q8_0 in this case, so the conversion errors out:

warning: blk.0.ffn_down_exps.weight - ncols 400 not divisible by 32 (required for type q8_0)
llama_model_quantize: failed to quantize: no tensor type fallback is defined for type q8_0

K-quants (Q4_K_M, Q6_K) use a 256-element super-block layout that handles 400 fine, hence they succeed. Q6_K is the practical substitute — its perplexity gap vs Q8_0 on Llama-3-8B reference is roughly +0.02 (negligible).

Measured throughput (RTX 3060 6 GB, num_ctx=8192, Ollama `/api/chat` with the Modelfile template override)

Measured on the orchestration_planner task (~1.7 KB system + short user query, ~190-token JSON output, temperature=0.0).

Quant	Eval rate	First-token latency	Notes
Q4_K_M	160 tok/s	~0.4 s (warm)	Recommended for interactive use
bf16	118 tok/s	~5 s (warm)	Same precision as the original safetensors; output literally matches ground truth

(Q6_K not benchmarked here; expected to fall between the two, closer to Q4_K_M. Numbers were obtained via /api/chat with the Go-template override — without that override Ollama strips role markers and decoding degenerates to < 1 tok/s, which is a chat-template issue, not a property of the weights.)

Evaluation

Judged by lightning-ai/gpt-oss-120b on 212 examples from TraceVerse-RL-Eval-Golden across 4 dimensions (1–10 scale). Detailed per-sample completions and scores are in TraceVerse-RL-Eval-Inference.

(Per-environment table is below in the auto-maintained benchmark block.)

Limitations

Domain-specialised. Best on structured trace-analysis prompts. Casual chat works but is not the trained distribution.
GGUF-embedded chat template does not work in Ollama — llama.cpp's Jinja parser cannot handle the source chat_template.jinja (macros, raise_exception, enable_thinking branches), so Ollama falls back to a bare {{ .Prompt }} template and /api/chat produces hallucinated output. Always use the Modelfile TEMPLATE override shown in the Ollama section above; with it the model emits correct training-distribution outputs (verified on the orchestration_planner task — output matches ground truth).
MTP heads not exported. Qwen3.5-MoE's multi-token-prediction layers were dropped during the source SFT; the GGUFs ship the main LM head only. Ollama / llama.cpp mainline do not consume MTP at inference, so this is irrelevant for typical usage. (Speculative-decoding-aware runtimes such as am17an/llama.cpp:mtp-clean would benefit if MTP were restored.)
Hallucination on out-of-domain technical questions is consistent with a small specialised model — verify generated factual claims independently.

Measuring throughput

The Ollama chat UI does not display tokens/s. To measure:

# CLI verbose mode
ollama run tracegenix-mini --verbose "your prompt"

# Via the API
curl -s http://localhost:11434/api/chat -d '{
  "model":"tracegenix-mini",
  "messages":[{"role":"user","content":"..."}],
  "stream":false
}' | jq '{eval_count, tokens_per_s: (.eval_count / (.eval_duration / 1e9))}'

# Or tail the server log
tail -f ~/.ollama/logs/server.log

Citation

@misc{tracegenix-mini-sft-clean-3ep,
  author = {Kshitij Thakkar},
  title  = {tracegenix-mini-sft-clean-3ep: a small Qwen3.5-MoE for OpenTelemetry trace analysis},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/kshitijthakkar/tracegenix-mini-sft-clean-3ep}}
}

Benchmark Results — TraceVerse-RL-Eval-Golden

Judged by lightning-ai/gpt-oss-120b on 212 examples across 4 dimensions (1-10 scale).

Overall	Accuracy	Completeness	Format	Grounding
4.25	3.86	4.29	5.23	4.11

Per-environment

Environment	N	Overall
orchestration_planner	23	7.89
cloud_incident_response	2	5.35
incident_ops	2	4.85
soc_triage_gym	3	4.47
tool_orchestrator	43	4.35
trace_classifier	60	3.89
report_generator	33	3.45
agent_scorer	30	3.38
deep_investigator	9	3.19
quality_evaluator	7	2.99

See full per-sample scores at kshitijthakkar/TraceVerse-RL-Eval-Inference.

Downloads last month: 507

Safetensors

Model size

1B params

Tensor type

BF16

Model tree for kshitijthakkar/tracegenix-mini-sft-clean-3ep

Base model

Qwen/Qwen3.5-0.8B-Base

Quantized

(29)

this model

kshitijthakkar
/

tracegenix-mini-sft-clean-3ep

tracegenix-mini-sft-clean-3ep

At a glance

Intended use

How to use

With Ollama (recommended for local use)

With `llama-server`

With `transformers` (bf16 safetensors)

GGUF quantizations

Why no Q8_0

Measured throughput (RTX 3060 6 GB, num_ctx=8192, Ollama `/api/chat` with the Modelfile template override)

Evaluation

Limitations

Measuring throughput

Citation

Benchmark Results — TraceVerse-RL-Eval-Golden

Per-environment

Model tree for kshitijthakkar/tracegenix-mini-sft-clean-3ep

Space using kshitijthakkar/tracegenix-mini-sft-clean-3ep 1

tracegenix-mini-sft-clean-3ep

At a glance

Intended use

How to use

With Ollama (recommended for local use)

With llama-server

With transformers (bf16 safetensors)

GGUF quantizations

Why no Q8_0

Measured throughput (RTX 3060 6 GB, num_ctx=8192, Ollama /api/chat with the Modelfile template override)

Evaluation

Limitations

Measuring throughput

Citation

Benchmark Results — TraceVerse-RL-Eval-Golden

Per-environment

Model tree for kshitijthakkar/tracegenix-mini-sft-clean-3ep

Space using kshitijthakkar/tracegenix-mini-sft-clean-3ep 1

With `llama-server`

With `transformers` (bf16 safetensors)

Measured throughput (RTX 3060 6 GB, num_ctx=8192, Ollama `/api/chat` with the Modelfile template override)