AnkitAI / Parable-Qwen3-4B-fable

Parable-Qwen3-4B-fable

First release of the Parable series: small models trained on real agent behavior, every release eval-gated before publish.

Parable-Qwen3-4B is a Qwen3-4B fine-tune trained on real Claude Fable 5 and GPT-5.5 agent traces: multi-step tool use, planning, and <think> reasoning captured from actual agent sessions, not synthetic Q&A. It cuts held-out test loss by 47% against the base under matched evaluation and reaches 0.782 token accuracy , within a point of models twice its size trained on similar traces.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AnkitAI/Parable-Qwen3-4B-fable"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Write a Python function that retries an HTTP request with exponential backoff."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=3000, temperature=0.3, do_sample=True)
text = tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True)

# reasoning model: final answer follows the </think> block
answer = text.split("</think>")[-1].strip()
print(answer)

GGUF quants for llama.cpp / Ollama / LM Studio: Parable-Qwen3-4B-fable-GGUF

Note: these weights are the F16 merge of a QLoRA adapter trained on the 4-bit base; quality is equivalent to the Q8 GGUF, published here for server stacks and further fine-tuning.

This is a reasoning model: output opens with a <think>...</think> block before the final answer. Strip it before showing responses to end users (llama.cpp's --jinja chat mode separates it automatically).

Sampling: temperature 0.3–0.7. Budget max_tokens generously ( ≥ 2500 ): like other trace-trained reasoning models, it thinks at length before answering, and a short budget can cut it off mid-thought.

Training data

• Glint-Research/Fable-5-traces : 4.4k real Claude Fable 5 coding-agent session traces with <think> reasoning and tool calls (AGPL-3.0)
• Roman1111111/gpt5.5-terminal : terminal-agent task solutions (MIT)

Every example passed a quality gate (schema validation, secrets scrub, length filtering) before training. QLoRA fine-tune via mlx-lm , quantized with llama.cpp .

Evaluation

Held-out test split, identical evaluation code for base and fine-tune (base measured through a zero-effect adapter for exact comparability):

Qualitative review (34 coding/terminal/debugging prompts, judged clean-and-correct): of the prompts that produced a final answer, 92% were correct . The remainder hit reasoning-budget cutoffs rather than wrong answers (23/34 overall with a 2,600-token budget; see guidance above).

Limitations

• Like other trace-trained reasoning models, it invests heavily in thinking. With tight token budgets it can spend the whole budget reasoning; budget ≥ 2500 tokens or retry at lower temperature if a response comes back empty.
• Tuned hard toward agentic coding behavior; that focus trades some general-knowledge breadth, as with any specialized fine-tune in this class.
• Verify critical output. Small models over-commit to plausible specifics; treat generated commands and code as drafts to review.
• Inherits Qwen3-4B's base limitations and knowledge cutoff.

Provenance & licensing

Model weights: Apache-2.0 (inherited from Qwen3-4B). Training data licenses: Fable-5-traces AGPL-3.0 , gpt5.5-terminal MIT . Because those traces originate from third-party assistants, the providers' terms may apply to downstream training and distillation. If you plan to build on this model commercially, confirm your use aligns with those terms.

Acknowledgements

• Glint-Research and Roman1111111 for the open trace datasets
• empero-ai , whose Qwable recipe this release follows
• Qwen team for the base model
• mlx-lm and llama.cpp