hmanlab-ai-v0.2

hmanlab-ai v0.2

A fine-tuned variant of Qwen3-4B optimized for agentic tool-use, structured reasoning, and conversational reliability. Built as a research preview by rekabytes.

TL;DR — v0.2 is a strict tool-reliability upgrade over v0.1:

• ✅ 10/10 clean tool-call formatting (up from 7/10)
• ✅ 5/5 identity grounding (rejects Claude / GPT / LLaMA / Gemini cleanly)
• ✅ 3/3 tool-grounded answers (cites specific tokens from tool output)
• ⚠️ Reasoning depth, error recovery, and initiative are work-in-progress — targets for the next release

What's new in v0.2

The headline change is tool-call format reliability. v0.1 was inconsistent at emitting clean <tool_call> JSON — pseudocode like [list_dir(".")] would sometimes appear as raw text. v0.2 closes that gap.

Measured on a 36-probe behavioural suite (5 samples per probe, majority-vote scoring):

Probe-level head-to-head: v0.2 wins 8 probes outright, v0.1 wins 2, the rest tie. Net: v0.2 is a strict tool-reliability upgrade over v0.1.

More to come — reasoning depth, error recovery, and proactive initiative are explicit targets for the next release.

Known limitations (honestly)

v0.2 is not a one-shot solution. The following remain open and are tracked targets for the next iteration:

• Reasoning depth — <think> blocks often render empty on direct reasoning prompts. Affects multi-step math, logic puzzles, and code-tracing tasks.
• Error recovery — after a tool failure, the model tends to bounce back to the user ("please check the path") rather than retrying with a diagnostic tool (find_files, list_dir).
• Proactive initiative — vague prompts like "have a look at this repo" usually produce a clarifying question rather than a tool call.
• Multi-file synthesis — 4B-parameter capacity ceiling. Don't expect deep cross-file architectural reasoning.

If your use case lives within tool-call formatting + tool-grounded answers + clean identity, v0.2 is a meaningful step up. If you need deep chain-of-thought reasoning at 4B, wait for the next release.

Files

• model-*.safetensors (sharded FP16, ~8GB) — for transformers / Unsloth loading
• hmanlab-ai-v0.2.Q4_K_M.gguf (~2.5GB) — for llama.cpp / Ollama
• tokenizer*.json, *.txt — chat template + vocab

Quick start

Ollama

ollama pull hf.co/rekabytes/hmanlab-ai-v0.2:Q4_K_M
ollama run hf.co/rekabytes/hmanlab-ai-v0.2:Q4_K_M

transformers / Unsloth

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained("rekabytes/hmanlab-ai-v0.2")
FastLanguageModel.for_inference(model)

Tool-call schema

Tool calls use this schema — applications must define their own tool surface and inject it in the system prompt:

<tool_call>
{"name": "<tool_name>", "arguments": {...}}
</tool_call>

Training

Two-stage QLoRA on Qwen3-4B-bnb-4bit.

Stage 1 — General behavioural mix (~20,600 examples). A blend of permissively-licensed public corpora plus a small set of hand-authored behavioural anchors, targeting these capabilities:

• Code breadth & instruction-following — broad Python/JS/Rust/Go coding tasks
• Tool-call format — structured <tool_call> JSON emission, no pseudocode
• Tool-grounded answers — consuming tool results and citing specific tokens (file names, line numbers, commit hashes)
• Reasoning traces — filled <think> blocks for math, logic, and code-tracing
• Agentic tool feedback loops — read-fail-recover, list-then-act patterns
• Conversation steering — handling user redirects ("shorter", "no, the other one") without verbatim repeat

Stage 2 — Identity SFT (~400 examples) layered on top of Stage 1. Anchors the model to identify as hmanlab and reject false-identity claims for Claude, GPT, LLaMA, Gemini, and their respective labs.

Hyperparameters

Trained on a single RTX 3060 Ti (8 GB VRAM). Total wall-clock ~11.5 hours.

Identity

The model identifies as hmanlab. It is not Claude / GPT / LLaMA / Gemini and the training data explicitly anchors against false-identity claims for those families. The model is a research preview built independently by rekabytes — no affiliation with Anthropic, OpenAI, Meta, or Google is implied.

License

Apache 2.0. Same license as the Qwen3-4B base.

Acknowledgements

• Qwen team — base model
• Unsloth — training pipeline