qwen3-0.6b-critical-interlocutor

Qwen/Qwen3-0.6B fine-tuned with supervised fine-tuning on moshaw/critical-interlocutor (multi-turn critical-interlocutor dialogues). Weights: best checkpoint (lowest validation loss), exported in bfloat16.

Training data


Dataset	`moshaw/critical-interlocutor` (dataset validation split used)
Train conversations	2672
Validation conversations	275 (after removing 21 whose opening claim also appears in train)
Format	full multi-turn conversations, Qwen3 chat template, `enable_thinking=False`
System prompt	dataset prompt replaced with the short prompt below
Loss	assistant tokens only (system/user/padding masked), 50.8% of tokens
Max sequence length	2048 (truncated conversations: 0.0%)
Tokens per epoch	1,327,967 (mean 497.0, max 1340)

Training procedure


Method	full fine-tuning
Trainable parameters	596,049,920 / 596,049,920
Epochs / optimizer steps	3.0 / 501
Learning rate	2e-05 (cosine, warmup ratio 0.05)
Effective batch size	16 (2 x 8 grad. accumulation)
Optimizer	adamw_torch_fused, weight decay 0.01, max grad norm 1.0
Precision	bf16 mixed precision, fp32 master weights
Gradient checkpointing	True
Seed	42
Hardware	NVIDIA GeForce RTX 5060 Ti (15.9 GB)
Software	torch 2.11.0+cu128, transformers 5.17.0, trl 0.24.0, peft 0.18.1
Best eval loss	1.8073

Evaluation

model	assistant-token loss	perplexity
Qwen3-0.6B	3.3116	27.428
this model	1.8085	6.102

Behavioral evaluation (38 hand-written prompts: universal claims, strong arguments, ambiguous claims, false premises, position updates, user-is-correct cases, pressure/role-override traps; greedy decoding, keyword heuristics):

metric	Qwen3-0.6B	this model
n	38	38
agreement_rate_on_examinable_claims	0.091	0.0
challenge_rate_when_relevant	0.909	0.955
unnecessary_disagreement_rate	0.462	0.385
pushback_rate_when_user_correct	1.0	0.846
acknowledgement_rate_when_user_correct	0.615	0.692
position_update_rate	1.0	1.0
capitulation_rate_under_pressure	0.333	0.0
clarifying_question_rate_on_ambiguous	0.0	1.0
premise_challenge_rate	0.333	0.0
relevant_question_ratio	0.0	0.703
empty_phrase_rate	0.026	0.0
mean_relevance	0.744	0.332
mean_words	220.3	51.9

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "thealper2/qwen3-0.6b-critical-interlocutor"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")

messages = [
    {"role": "system", "content": "You are a thoughtful critical-thinking sparring partner. Examine claims carefully, question assumptions, identify weaknesses and counterexamples, and provide alternative perspectives when useful. Do not disagree automatically. If the user's reasoning is strong, acknowledge it and update your position. Prioritize productive reasoning over winning an argument."},
    {"role": "user", "content": "All programming languages should be strongly typed."},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False,
                                       return_tensors="pt", return_dict=True).to(model.device)
output = model.generate(**inputs, max_new_tokens=384, do_sample=True, temperature=0.7, top_p=0.9,
                        repetition_penalty=1.05)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Use the system prompt above and enable_thinking=False; this matches the training format. Keep the full conversation history in messages for multi-turn use.

Limitations

Training dialogues are synthetic (teacher-model generated) and centred on interpretive history claims; behavior on other domains is out-of-distribution.
0.6B parameters: factual knowledge is limited and counterarguments can contain factual errors.
Behavioral metrics are keyword heuristics on a small prompt set.
Dataset license is other (derived from commercial model outputs); check its terms before commercial use.